2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
27 #include "anv_private.h"
28 #include "vk_format_info.h"
30 #include "util/fast_idiv_by_const.h"
32 #include "common/gen_aux_map.h"
33 #include "common/gen_l3_config.h"
34 #include "genxml/gen_macros.h"
35 #include "genxml/genX_pack.h"
37 /* We reserve GPR 14 and 15 for conditional rendering */
38 #define GEN_MI_BUILDER_NUM_ALLOC_GPRS 14
39 #define __gen_get_batch_dwords anv_batch_emit_dwords
40 #define __gen_address_offset anv_address_add
41 #include "common/gen_mi_builder.h"
43 static void genX(flush_pipeline_select
)(struct anv_cmd_buffer
*cmd_buffer
,
47 emit_lri(struct anv_batch
*batch
, uint32_t reg
, uint32_t imm
)
49 anv_batch_emit(batch
, GENX(MI_LOAD_REGISTER_IMM
), lri
) {
50 lri
.RegisterOffset
= reg
;
56 genX(cmd_buffer_emit_state_base_address
)(struct anv_cmd_buffer
*cmd_buffer
)
58 struct anv_device
*device
= cmd_buffer
->device
;
59 UNUSED
const struct gen_device_info
*devinfo
= &device
->info
;
60 uint32_t mocs
= device
->isl_dev
.mocs
.internal
;
62 /* If we are emitting a new state base address we probably need to re-emit
65 cmd_buffer
->state
.descriptors_dirty
|= ~0;
67 /* Emit a render target cache flush.
69 * This isn't documented anywhere in the PRM. However, it seems to be
70 * necessary prior to changing the surface state base adress. Without
71 * this, we get GPU hangs when using multi-level command buffers which
72 * clear depth, reset state base address, and then go render stuff.
74 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
75 pc
.DCFlushEnable
= true;
76 pc
.RenderTargetCacheFlushEnable
= true;
77 pc
.CommandStreamerStallEnable
= true;
79 pc
.TileCacheFlushEnable
= true;
82 /* GEN:BUG:1606662791:
84 * Software must program PIPE_CONTROL command with "HDC Pipeline
85 * Flush" prior to programming of the below two non-pipeline state :
86 * * STATE_BASE_ADDRESS
87 * * 3DSTATE_BINDING_TABLE_POOL_ALLOC
89 if (devinfo
->revision
== 0 /* A0 */)
90 pc
.HDCPipelineFlushEnable
= true;
95 /* GEN:BUG:1607854226:
97 * Workaround the non pipelined state not applying in MEDIA/GPGPU pipeline
98 * mode by putting the pipeline temporarily in 3D mode.
100 uint32_t gen12_wa_pipeline
= cmd_buffer
->state
.current_pipeline
;
101 genX(flush_pipeline_select_3d
)(cmd_buffer
);
104 anv_batch_emit(&cmd_buffer
->batch
, GENX(STATE_BASE_ADDRESS
), sba
) {
105 sba
.GeneralStateBaseAddress
= (struct anv_address
) { NULL
, 0 };
106 sba
.GeneralStateMOCS
= mocs
;
107 sba
.GeneralStateBaseAddressModifyEnable
= true;
109 sba
.StatelessDataPortAccessMOCS
= mocs
;
111 sba
.SurfaceStateBaseAddress
=
112 anv_cmd_buffer_surface_base_address(cmd_buffer
);
113 sba
.SurfaceStateMOCS
= mocs
;
114 sba
.SurfaceStateBaseAddressModifyEnable
= true;
116 sba
.DynamicStateBaseAddress
=
117 (struct anv_address
) { device
->dynamic_state_pool
.block_pool
.bo
, 0 };
118 sba
.DynamicStateMOCS
= mocs
;
119 sba
.DynamicStateBaseAddressModifyEnable
= true;
121 sba
.IndirectObjectBaseAddress
= (struct anv_address
) { NULL
, 0 };
122 sba
.IndirectObjectMOCS
= mocs
;
123 sba
.IndirectObjectBaseAddressModifyEnable
= true;
125 sba
.InstructionBaseAddress
=
126 (struct anv_address
) { device
->instruction_state_pool
.block_pool
.bo
, 0 };
127 sba
.InstructionMOCS
= mocs
;
128 sba
.InstructionBaseAddressModifyEnable
= true;
131 /* Broadwell requires that we specify a buffer size for a bunch of
132 * these fields. However, since we will be growing the BO's live, we
133 * just set them all to the maximum.
135 sba
.GeneralStateBufferSize
= 0xfffff;
136 sba
.IndirectObjectBufferSize
= 0xfffff;
137 if (device
->physical
->use_softpin
) {
138 /* With softpin, we use fixed addresses so we actually know how big
139 * our base addresses are.
141 sba
.DynamicStateBufferSize
= DYNAMIC_STATE_POOL_SIZE
/ 4096;
142 sba
.InstructionBufferSize
= INSTRUCTION_STATE_POOL_SIZE
/ 4096;
144 sba
.DynamicStateBufferSize
= 0xfffff;
145 sba
.InstructionBufferSize
= 0xfffff;
147 sba
.GeneralStateBufferSizeModifyEnable
= true;
148 sba
.IndirectObjectBufferSizeModifyEnable
= true;
149 sba
.DynamicStateBufferSizeModifyEnable
= true;
150 sba
.InstructionBuffersizeModifyEnable
= true;
152 /* On gen7, we have upper bounds instead. According to the docs,
153 * setting an upper bound of zero means that no bounds checking is
154 * performed so, in theory, we should be able to leave them zero.
155 * However, border color is broken and the GPU bounds-checks anyway.
156 * To avoid this and other potential problems, we may as well set it
159 sba
.GeneralStateAccessUpperBound
=
160 (struct anv_address
) { .bo
= NULL
, .offset
= 0xfffff000 };
161 sba
.GeneralStateAccessUpperBoundModifyEnable
= true;
162 sba
.DynamicStateAccessUpperBound
=
163 (struct anv_address
) { .bo
= NULL
, .offset
= 0xfffff000 };
164 sba
.DynamicStateAccessUpperBoundModifyEnable
= true;
165 sba
.InstructionAccessUpperBound
=
166 (struct anv_address
) { .bo
= NULL
, .offset
= 0xfffff000 };
167 sba
.InstructionAccessUpperBoundModifyEnable
= true;
170 if (cmd_buffer
->device
->physical
->use_softpin
) {
171 sba
.BindlessSurfaceStateBaseAddress
= (struct anv_address
) {
172 .bo
= device
->surface_state_pool
.block_pool
.bo
,
175 sba
.BindlessSurfaceStateSize
= (1 << 20) - 1;
177 sba
.BindlessSurfaceStateBaseAddress
= ANV_NULL_ADDRESS
;
178 sba
.BindlessSurfaceStateSize
= 0;
180 sba
.BindlessSurfaceStateMOCS
= mocs
;
181 sba
.BindlessSurfaceStateBaseAddressModifyEnable
= true;
184 sba
.BindlessSamplerStateBaseAddress
= (struct anv_address
) { NULL
, 0 };
185 sba
.BindlessSamplerStateMOCS
= mocs
;
186 sba
.BindlessSamplerStateBaseAddressModifyEnable
= true;
187 sba
.BindlessSamplerStateBufferSize
= 0;
192 /* GEN:BUG:1607854226:
194 * Put the pipeline back into its current mode.
196 if (gen12_wa_pipeline
!= UINT32_MAX
)
197 genX(flush_pipeline_select
)(cmd_buffer
, gen12_wa_pipeline
);
200 /* After re-setting the surface state base address, we have to do some
201 * cache flusing so that the sampler engine will pick up the new
202 * SURFACE_STATE objects and binding tables. From the Broadwell PRM,
203 * Shared Function > 3D Sampler > State > State Caching (page 96):
205 * Coherency with system memory in the state cache, like the texture
206 * cache is handled partially by software. It is expected that the
207 * command stream or shader will issue Cache Flush operation or
208 * Cache_Flush sampler message to ensure that the L1 cache remains
209 * coherent with system memory.
213 * Whenever the value of the Dynamic_State_Base_Addr,
214 * Surface_State_Base_Addr are altered, the L1 state cache must be
215 * invalidated to ensure the new surface or sampler state is fetched
216 * from system memory.
218 * The PIPE_CONTROL command has a "State Cache Invalidation Enable" bit
219 * which, according the PIPE_CONTROL instruction documentation in the
222 * Setting this bit is independent of any other bit in this packet.
223 * This bit controls the invalidation of the L1 and L2 state caches
224 * at the top of the pipe i.e. at the parsing time.
226 * Unfortunately, experimentation seems to indicate that state cache
227 * invalidation through a PIPE_CONTROL does nothing whatsoever in
228 * regards to surface state and binding tables. In stead, it seems that
229 * invalidating the texture cache is what is actually needed.
231 * XXX: As far as we have been able to determine through
232 * experimentation, shows that flush the texture cache appears to be
233 * sufficient. The theory here is that all of the sampling/rendering
234 * units cache the binding table in the texture cache. However, we have
235 * yet to be able to actually confirm this.
237 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
238 pc
.TextureCacheInvalidationEnable
= true;
239 pc
.ConstantCacheInvalidationEnable
= true;
240 pc
.StateCacheInvalidationEnable
= true;
245 add_surface_reloc(struct anv_cmd_buffer
*cmd_buffer
,
246 struct anv_state state
, struct anv_address addr
)
248 const struct isl_device
*isl_dev
= &cmd_buffer
->device
->isl_dev
;
251 anv_reloc_list_add(&cmd_buffer
->surface_relocs
, &cmd_buffer
->pool
->alloc
,
252 state
.offset
+ isl_dev
->ss
.addr_offset
,
253 addr
.bo
, addr
.offset
, NULL
);
254 if (result
!= VK_SUCCESS
)
255 anv_batch_set_error(&cmd_buffer
->batch
, result
);
259 add_surface_state_relocs(struct anv_cmd_buffer
*cmd_buffer
,
260 struct anv_surface_state state
)
262 const struct isl_device
*isl_dev
= &cmd_buffer
->device
->isl_dev
;
264 assert(!anv_address_is_null(state
.address
));
265 add_surface_reloc(cmd_buffer
, state
.state
, state
.address
);
267 if (!anv_address_is_null(state
.aux_address
)) {
269 anv_reloc_list_add(&cmd_buffer
->surface_relocs
,
270 &cmd_buffer
->pool
->alloc
,
271 state
.state
.offset
+ isl_dev
->ss
.aux_addr_offset
,
272 state
.aux_address
.bo
,
273 state
.aux_address
.offset
,
275 if (result
!= VK_SUCCESS
)
276 anv_batch_set_error(&cmd_buffer
->batch
, result
);
279 if (!anv_address_is_null(state
.clear_address
)) {
281 anv_reloc_list_add(&cmd_buffer
->surface_relocs
,
282 &cmd_buffer
->pool
->alloc
,
284 isl_dev
->ss
.clear_color_state_offset
,
285 state
.clear_address
.bo
,
286 state
.clear_address
.offset
,
288 if (result
!= VK_SUCCESS
)
289 anv_batch_set_error(&cmd_buffer
->batch
, result
);
294 color_attachment_compute_aux_usage(struct anv_device
* device
,
295 struct anv_cmd_state
* cmd_state
,
296 uint32_t att
, VkRect2D render_area
,
297 union isl_color_value
*fast_clear_color
)
299 struct anv_attachment_state
*att_state
= &cmd_state
->attachments
[att
];
300 struct anv_image_view
*iview
= cmd_state
->attachments
[att
].image_view
;
302 assert(iview
->n_planes
== 1);
304 if (iview
->planes
[0].isl
.base_array_layer
>=
305 anv_image_aux_layers(iview
->image
, VK_IMAGE_ASPECT_COLOR_BIT
,
306 iview
->planes
[0].isl
.base_level
)) {
307 /* There is no aux buffer which corresponds to the level and layer(s)
310 att_state
->aux_usage
= ISL_AUX_USAGE_NONE
;
311 att_state
->input_aux_usage
= ISL_AUX_USAGE_NONE
;
312 att_state
->fast_clear
= false;
316 att_state
->aux_usage
=
317 anv_layout_to_aux_usage(&device
->info
, iview
->image
,
318 VK_IMAGE_ASPECT_COLOR_BIT
,
319 VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT
,
320 VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL
);
322 /* If we don't have aux, then we should have returned early in the layer
323 * check above. If we got here, we must have something.
325 assert(att_state
->aux_usage
!= ISL_AUX_USAGE_NONE
);
327 if (att_state
->aux_usage
== ISL_AUX_USAGE_CCS_E
||
328 att_state
->aux_usage
== ISL_AUX_USAGE_MCS
) {
329 att_state
->input_aux_usage
= att_state
->aux_usage
;
331 /* From the Sky Lake PRM, RENDER_SURFACE_STATE::AuxiliarySurfaceMode:
333 * "If Number of Multisamples is MULTISAMPLECOUNT_1, AUX_CCS_D
334 * setting is only allowed if Surface Format supported for Fast
335 * Clear. In addition, if the surface is bound to the sampling
336 * engine, Surface Format must be supported for Render Target
337 * Compression for surfaces bound to the sampling engine."
339 * In other words, we can only sample from a fast-cleared image if it
340 * also supports color compression.
342 if (isl_format_supports_ccs_e(&device
->info
, iview
->planes
[0].isl
.format
) &&
343 isl_format_supports_ccs_d(&device
->info
, iview
->planes
[0].isl
.format
)) {
344 att_state
->input_aux_usage
= ISL_AUX_USAGE_CCS_D
;
346 /* While fast-clear resolves and partial resolves are fairly cheap in the
347 * case where you render to most of the pixels, full resolves are not
348 * because they potentially involve reading and writing the entire
349 * framebuffer. If we can't texture with CCS_E, we should leave it off and
350 * limit ourselves to fast clears.
352 if (cmd_state
->pass
->attachments
[att
].first_subpass_layout
==
353 VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL
) {
354 anv_perf_warn(device
, iview
->image
,
355 "Not temporarily enabling CCS_E.");
358 att_state
->input_aux_usage
= ISL_AUX_USAGE_NONE
;
362 assert(iview
->image
->planes
[0].aux_surface
.isl
.usage
&
363 (ISL_SURF_USAGE_CCS_BIT
| ISL_SURF_USAGE_MCS_BIT
));
365 union isl_color_value clear_color
= {};
366 anv_clear_color_from_att_state(&clear_color
, att_state
, iview
);
368 att_state
->clear_color_is_zero_one
=
369 isl_color_value_is_zero_one(clear_color
, iview
->planes
[0].isl
.format
);
370 att_state
->clear_color_is_zero
=
371 isl_color_value_is_zero(clear_color
, iview
->planes
[0].isl
.format
);
373 if (att_state
->pending_clear_aspects
== VK_IMAGE_ASPECT_COLOR_BIT
) {
374 /* Start by getting the fast clear type. We use the first subpass
375 * layout here because we don't want to fast-clear if the first subpass
376 * to use the attachment can't handle fast-clears.
378 enum anv_fast_clear_type fast_clear_type
=
379 anv_layout_to_fast_clear_type(&device
->info
, iview
->image
,
380 VK_IMAGE_ASPECT_COLOR_BIT
,
381 cmd_state
->pass
->attachments
[att
].first_subpass_layout
);
382 switch (fast_clear_type
) {
383 case ANV_FAST_CLEAR_NONE
:
384 att_state
->fast_clear
= false;
386 case ANV_FAST_CLEAR_DEFAULT_VALUE
:
387 att_state
->fast_clear
= att_state
->clear_color_is_zero
;
389 case ANV_FAST_CLEAR_ANY
:
390 att_state
->fast_clear
= true;
394 /* Potentially, we could do partial fast-clears but doing so has crazy
395 * alignment restrictions. It's easier to just restrict to full size
396 * fast clears for now.
398 if (render_area
.offset
.x
!= 0 ||
399 render_area
.offset
.y
!= 0 ||
400 render_area
.extent
.width
!= iview
->extent
.width
||
401 render_area
.extent
.height
!= iview
->extent
.height
)
402 att_state
->fast_clear
= false;
404 /* On Broadwell and earlier, we can only handle 0/1 clear colors */
405 if (GEN_GEN
<= 8 && !att_state
->clear_color_is_zero_one
)
406 att_state
->fast_clear
= false;
408 /* We only allow fast clears to the first slice of an image (level 0,
409 * layer 0) and only for the entire slice. This guarantees us that, at
410 * any given time, there is only one clear color on any given image at
411 * any given time. At the time of our testing (Jan 17, 2018), there
412 * were no known applications which would benefit from fast-clearing
413 * more than just the first slice.
415 if (att_state
->fast_clear
&&
416 (iview
->planes
[0].isl
.base_level
> 0 ||
417 iview
->planes
[0].isl
.base_array_layer
> 0)) {
418 anv_perf_warn(device
, iview
->image
,
419 "Rendering with multi-lod or multi-layer framebuffer "
420 "with LOAD_OP_LOAD and baseMipLevel > 0 or "
421 "baseArrayLayer > 0. Not fast clearing.");
422 att_state
->fast_clear
= false;
423 } else if (att_state
->fast_clear
&& cmd_state
->framebuffer
->layers
> 1) {
424 anv_perf_warn(device
, iview
->image
,
425 "Rendering to a multi-layer framebuffer with "
426 "LOAD_OP_CLEAR. Only fast-clearing the first slice");
429 if (att_state
->fast_clear
)
430 *fast_clear_color
= clear_color
;
432 att_state
->fast_clear
= false;
437 depth_stencil_attachment_compute_aux_usage(struct anv_device
*device
,
438 struct anv_cmd_state
*cmd_state
,
439 uint32_t att
, VkRect2D render_area
)
441 struct anv_render_pass_attachment
*pass_att
=
442 &cmd_state
->pass
->attachments
[att
];
443 struct anv_attachment_state
*att_state
= &cmd_state
->attachments
[att
];
444 struct anv_image_view
*iview
= cmd_state
->attachments
[att
].image_view
;
446 /* These will be initialized after the first subpass transition. */
447 att_state
->aux_usage
= ISL_AUX_USAGE_NONE
;
448 att_state
->input_aux_usage
= ISL_AUX_USAGE_NONE
;
450 /* This is unused for depth/stencil but valgrind complains if it
453 att_state
->clear_color_is_zero_one
= false;
456 /* We don't do any HiZ or depth fast-clears on gen7 yet */
457 att_state
->fast_clear
= false;
461 if (!(att_state
->pending_clear_aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
)) {
462 /* If we're just clearing stencil, we can always HiZ clear */
463 att_state
->fast_clear
= true;
467 /* Default to false for now */
468 att_state
->fast_clear
= false;
470 /* We must have depth in order to have HiZ */
471 if (!(iview
->image
->aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
))
474 const enum isl_aux_usage first_subpass_aux_usage
=
475 anv_layout_to_aux_usage(&device
->info
, iview
->image
,
476 VK_IMAGE_ASPECT_DEPTH_BIT
,
477 VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT
,
478 pass_att
->first_subpass_layout
);
479 if (!blorp_can_hiz_clear_depth(&device
->info
,
480 &iview
->image
->planes
[0].surface
.isl
,
481 first_subpass_aux_usage
,
482 iview
->planes
[0].isl
.base_level
,
483 iview
->planes
[0].isl
.base_array_layer
,
484 render_area
.offset
.x
,
485 render_area
.offset
.y
,
486 render_area
.offset
.x
+
487 render_area
.extent
.width
,
488 render_area
.offset
.y
+
489 render_area
.extent
.height
))
492 if (att_state
->clear_value
.depthStencil
.depth
!= ANV_HZ_FC_VAL
)
495 if (GEN_GEN
== 8 && anv_can_sample_with_hiz(&device
->info
, iview
->image
)) {
496 /* Only gen9+ supports returning ANV_HZ_FC_VAL when sampling a
497 * fast-cleared portion of a HiZ buffer. Testing has revealed that Gen8
498 * only supports returning 0.0f. Gens prior to gen8 do not support this
504 /* If we got here, then we can fast clear */
505 att_state
->fast_clear
= true;
509 need_input_attachment_state(const struct anv_render_pass_attachment
*att
)
511 if (!(att
->usage
& VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT
))
514 /* We only allocate input attachment states for color surfaces. Compression
515 * is not yet enabled for depth textures and stencil doesn't allow
516 * compression so we can just use the texture surface state from the view.
518 return vk_format_is_color(att
->format
);
521 /* Transitions a HiZ-enabled depth buffer from one layout to another. Unless
522 * the initial layout is undefined, the HiZ buffer and depth buffer will
523 * represent the same data at the end of this operation.
526 transition_depth_buffer(struct anv_cmd_buffer
*cmd_buffer
,
527 const struct anv_image
*image
,
528 VkImageLayout initial_layout
,
529 VkImageLayout final_layout
)
531 uint32_t depth_plane
=
532 anv_image_aspect_to_plane(image
->aspects
, VK_IMAGE_ASPECT_DEPTH_BIT
);
533 if (image
->planes
[depth_plane
].aux_usage
== ISL_AUX_USAGE_NONE
)
536 const enum isl_aux_state initial_state
=
537 anv_layout_to_aux_state(&cmd_buffer
->device
->info
, image
,
538 VK_IMAGE_ASPECT_DEPTH_BIT
,
540 const enum isl_aux_state final_state
=
541 anv_layout_to_aux_state(&cmd_buffer
->device
->info
, image
,
542 VK_IMAGE_ASPECT_DEPTH_BIT
,
545 const bool initial_depth_valid
=
546 isl_aux_state_has_valid_primary(initial_state
);
547 const bool initial_hiz_valid
=
548 isl_aux_state_has_valid_aux(initial_state
);
549 const bool final_needs_depth
=
550 isl_aux_state_has_valid_primary(final_state
);
551 const bool final_needs_hiz
=
552 isl_aux_state_has_valid_aux(final_state
);
554 /* Getting into the pass-through state for Depth is tricky and involves
555 * both a resolve and an ambiguate. We don't handle that state right now
556 * as anv_layout_to_aux_state never returns it.
558 assert(final_state
!= ISL_AUX_STATE_PASS_THROUGH
);
560 if (final_needs_depth
&& !initial_depth_valid
) {
561 assert(initial_hiz_valid
);
562 anv_image_hiz_op(cmd_buffer
, image
, VK_IMAGE_ASPECT_DEPTH_BIT
,
563 0, 0, 1, ISL_AUX_OP_FULL_RESOLVE
);
564 } else if (final_needs_hiz
&& !initial_hiz_valid
) {
565 assert(initial_depth_valid
);
566 anv_image_hiz_op(cmd_buffer
, image
, VK_IMAGE_ASPECT_DEPTH_BIT
,
567 0, 0, 1, ISL_AUX_OP_AMBIGUATE
);
572 vk_image_layout_stencil_write_optimal(VkImageLayout layout
)
574 return layout
== VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL
||
575 layout
== VK_IMAGE_LAYOUT_DEPTH_READ_ONLY_STENCIL_ATTACHMENT_OPTIMAL
||
576 layout
== VK_IMAGE_LAYOUT_STENCIL_ATTACHMENT_OPTIMAL_KHR
;
579 /* Transitions a HiZ-enabled depth buffer from one layout to another. Unless
580 * the initial layout is undefined, the HiZ buffer and depth buffer will
581 * represent the same data at the end of this operation.
584 transition_stencil_buffer(struct anv_cmd_buffer
*cmd_buffer
,
585 const struct anv_image
*image
,
586 uint32_t base_level
, uint32_t level_count
,
587 uint32_t base_layer
, uint32_t layer_count
,
588 VkImageLayout initial_layout
,
589 VkImageLayout final_layout
)
592 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
593 VK_IMAGE_ASPECT_STENCIL_BIT
);
595 /* On gen7, we have to store a texturable version of the stencil buffer in
596 * a shadow whenever VK_IMAGE_USAGE_SAMPLED_BIT is set and copy back and
597 * forth at strategic points. Stencil writes are only allowed in following
600 * - VK_IMAGE_LAYOUT_GENERAL
601 * - VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
602 * - VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL
603 * - VK_IMAGE_LAYOUT_DEPTH_READ_ONLY_STENCIL_ATTACHMENT_OPTIMAL
604 * - VK_IMAGE_LAYOUT_STENCIL_ATTACHMENT_OPTIMAL_KHR
606 * For general, we have no nice opportunity to transition so we do the copy
607 * to the shadow unconditionally at the end of the subpass. For transfer
608 * destinations, we can update it as part of the transfer op. For the other
609 * layouts, we delay the copy until a transition into some other layout.
611 if (image
->planes
[plane
].shadow_surface
.isl
.size_B
> 0 &&
612 vk_image_layout_stencil_write_optimal(initial_layout
) &&
613 !vk_image_layout_stencil_write_optimal(final_layout
)) {
614 anv_image_copy_to_shadow(cmd_buffer
, image
,
615 VK_IMAGE_ASPECT_STENCIL_BIT
,
616 base_level
, level_count
,
617 base_layer
, layer_count
);
619 #endif /* GEN_GEN == 7 */
622 #define MI_PREDICATE_SRC0 0x2400
623 #define MI_PREDICATE_SRC1 0x2408
624 #define MI_PREDICATE_RESULT 0x2418
627 set_image_compressed_bit(struct anv_cmd_buffer
*cmd_buffer
,
628 const struct anv_image
*image
,
629 VkImageAspectFlagBits aspect
,
631 uint32_t base_layer
, uint32_t layer_count
,
634 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
, aspect
);
636 /* We only have compression tracking for CCS_E */
637 if (image
->planes
[plane
].aux_usage
!= ISL_AUX_USAGE_CCS_E
)
640 for (uint32_t a
= 0; a
< layer_count
; a
++) {
641 uint32_t layer
= base_layer
+ a
;
642 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_STORE_DATA_IMM
), sdi
) {
643 sdi
.Address
= anv_image_get_compression_state_addr(cmd_buffer
->device
,
646 sdi
.ImmediateData
= compressed
? UINT32_MAX
: 0;
652 set_image_fast_clear_state(struct anv_cmd_buffer
*cmd_buffer
,
653 const struct anv_image
*image
,
654 VkImageAspectFlagBits aspect
,
655 enum anv_fast_clear_type fast_clear
)
657 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_STORE_DATA_IMM
), sdi
) {
658 sdi
.Address
= anv_image_get_fast_clear_type_addr(cmd_buffer
->device
,
660 sdi
.ImmediateData
= fast_clear
;
663 /* Whenever we have fast-clear, we consider that slice to be compressed.
664 * This makes building predicates much easier.
666 if (fast_clear
!= ANV_FAST_CLEAR_NONE
)
667 set_image_compressed_bit(cmd_buffer
, image
, aspect
, 0, 0, 1, true);
670 /* This is only really practical on haswell and above because it requires
671 * MI math in order to get it correct.
673 #if GEN_GEN >= 8 || GEN_IS_HASWELL
675 anv_cmd_compute_resolve_predicate(struct anv_cmd_buffer
*cmd_buffer
,
676 const struct anv_image
*image
,
677 VkImageAspectFlagBits aspect
,
678 uint32_t level
, uint32_t array_layer
,
679 enum isl_aux_op resolve_op
,
680 enum anv_fast_clear_type fast_clear_supported
)
682 struct gen_mi_builder b
;
683 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
685 const struct gen_mi_value fast_clear_type
=
686 gen_mi_mem32(anv_image_get_fast_clear_type_addr(cmd_buffer
->device
,
689 if (resolve_op
== ISL_AUX_OP_FULL_RESOLVE
) {
690 /* In this case, we're doing a full resolve which means we want the
691 * resolve to happen if any compression (including fast-clears) is
694 * In order to simplify the logic a bit, we make the assumption that,
695 * if the first slice has been fast-cleared, it is also marked as
696 * compressed. See also set_image_fast_clear_state.
698 const struct gen_mi_value compression_state
=
699 gen_mi_mem32(anv_image_get_compression_state_addr(cmd_buffer
->device
,
701 level
, array_layer
));
702 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC0
),
704 gen_mi_store(&b
, compression_state
, gen_mi_imm(0));
706 if (level
== 0 && array_layer
== 0) {
707 /* If the predicate is true, we want to write 0 to the fast clear type
708 * and, if it's false, leave it alone. We can do this by writing
710 * clear_type = clear_type & ~predicate;
712 struct gen_mi_value new_fast_clear_type
=
713 gen_mi_iand(&b
, fast_clear_type
,
714 gen_mi_inot(&b
, gen_mi_reg64(MI_PREDICATE_SRC0
)));
715 gen_mi_store(&b
, fast_clear_type
, new_fast_clear_type
);
717 } else if (level
== 0 && array_layer
== 0) {
718 /* In this case, we are doing a partial resolve to get rid of fast-clear
719 * colors. We don't care about the compression state but we do care
720 * about how much fast clear is allowed by the final layout.
722 assert(resolve_op
== ISL_AUX_OP_PARTIAL_RESOLVE
);
723 assert(fast_clear_supported
< ANV_FAST_CLEAR_ANY
);
725 /* We need to compute (fast_clear_supported < image->fast_clear) */
726 struct gen_mi_value pred
=
727 gen_mi_ult(&b
, gen_mi_imm(fast_clear_supported
), fast_clear_type
);
728 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC0
),
729 gen_mi_value_ref(&b
, pred
));
731 /* If the predicate is true, we want to write 0 to the fast clear type
732 * and, if it's false, leave it alone. We can do this by writing
734 * clear_type = clear_type & ~predicate;
736 struct gen_mi_value new_fast_clear_type
=
737 gen_mi_iand(&b
, fast_clear_type
, gen_mi_inot(&b
, pred
));
738 gen_mi_store(&b
, fast_clear_type
, new_fast_clear_type
);
740 /* In this case, we're trying to do a partial resolve on a slice that
741 * doesn't have clear color. There's nothing to do.
743 assert(resolve_op
== ISL_AUX_OP_PARTIAL_RESOLVE
);
747 /* Set src1 to 0 and use a != condition */
748 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC1
), gen_mi_imm(0));
750 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_PREDICATE
), mip
) {
751 mip
.LoadOperation
= LOAD_LOADINV
;
752 mip
.CombineOperation
= COMBINE_SET
;
753 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
756 #endif /* GEN_GEN >= 8 || GEN_IS_HASWELL */
760 anv_cmd_simple_resolve_predicate(struct anv_cmd_buffer
*cmd_buffer
,
761 const struct anv_image
*image
,
762 VkImageAspectFlagBits aspect
,
763 uint32_t level
, uint32_t array_layer
,
764 enum isl_aux_op resolve_op
,
765 enum anv_fast_clear_type fast_clear_supported
)
767 struct gen_mi_builder b
;
768 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
770 struct gen_mi_value fast_clear_type_mem
=
771 gen_mi_mem32(anv_image_get_fast_clear_type_addr(cmd_buffer
->device
,
774 /* This only works for partial resolves and only when the clear color is
775 * all or nothing. On the upside, this emits less command streamer code
776 * and works on Ivybridge and Bay Trail.
778 assert(resolve_op
== ISL_AUX_OP_PARTIAL_RESOLVE
);
779 assert(fast_clear_supported
!= ANV_FAST_CLEAR_ANY
);
781 /* We don't support fast clears on anything other than the first slice. */
782 if (level
> 0 || array_layer
> 0)
785 /* On gen8, we don't have a concept of default clear colors because we
786 * can't sample from CCS surfaces. It's enough to just load the fast clear
787 * state into the predicate register.
789 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC0
), fast_clear_type_mem
);
790 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC1
), gen_mi_imm(0));
791 gen_mi_store(&b
, fast_clear_type_mem
, gen_mi_imm(0));
793 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_PREDICATE
), mip
) {
794 mip
.LoadOperation
= LOAD_LOADINV
;
795 mip
.CombineOperation
= COMBINE_SET
;
796 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
799 #endif /* GEN_GEN <= 8 */
802 anv_cmd_predicated_ccs_resolve(struct anv_cmd_buffer
*cmd_buffer
,
803 const struct anv_image
*image
,
804 enum isl_format format
,
805 VkImageAspectFlagBits aspect
,
806 uint32_t level
, uint32_t array_layer
,
807 enum isl_aux_op resolve_op
,
808 enum anv_fast_clear_type fast_clear_supported
)
810 const uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
, aspect
);
813 anv_cmd_compute_resolve_predicate(cmd_buffer
, image
,
814 aspect
, level
, array_layer
,
815 resolve_op
, fast_clear_supported
);
816 #else /* GEN_GEN <= 8 */
817 anv_cmd_simple_resolve_predicate(cmd_buffer
, image
,
818 aspect
, level
, array_layer
,
819 resolve_op
, fast_clear_supported
);
822 /* CCS_D only supports full resolves and BLORP will assert on us if we try
823 * to do a partial resolve on a CCS_D surface.
825 if (resolve_op
== ISL_AUX_OP_PARTIAL_RESOLVE
&&
826 image
->planes
[plane
].aux_usage
== ISL_AUX_USAGE_CCS_D
)
827 resolve_op
= ISL_AUX_OP_FULL_RESOLVE
;
829 anv_image_ccs_op(cmd_buffer
, image
, format
, aspect
, level
,
830 array_layer
, 1, resolve_op
, NULL
, true);
834 anv_cmd_predicated_mcs_resolve(struct anv_cmd_buffer
*cmd_buffer
,
835 const struct anv_image
*image
,
836 enum isl_format format
,
837 VkImageAspectFlagBits aspect
,
838 uint32_t array_layer
,
839 enum isl_aux_op resolve_op
,
840 enum anv_fast_clear_type fast_clear_supported
)
842 assert(aspect
== VK_IMAGE_ASPECT_COLOR_BIT
);
843 assert(resolve_op
== ISL_AUX_OP_PARTIAL_RESOLVE
);
845 #if GEN_GEN >= 8 || GEN_IS_HASWELL
846 anv_cmd_compute_resolve_predicate(cmd_buffer
, image
,
847 aspect
, 0, array_layer
,
848 resolve_op
, fast_clear_supported
);
850 anv_image_mcs_op(cmd_buffer
, image
, format
, aspect
,
851 array_layer
, 1, resolve_op
, NULL
, true);
853 unreachable("MCS resolves are unsupported on Ivybridge and Bay Trail");
858 genX(cmd_buffer_mark_image_written
)(struct anv_cmd_buffer
*cmd_buffer
,
859 const struct anv_image
*image
,
860 VkImageAspectFlagBits aspect
,
861 enum isl_aux_usage aux_usage
,
864 uint32_t layer_count
)
866 /* The aspect must be exactly one of the image aspects. */
867 assert(util_bitcount(aspect
) == 1 && (aspect
& image
->aspects
));
869 /* The only compression types with more than just fast-clears are MCS,
870 * CCS_E, and HiZ. With HiZ we just trust the layout and don't actually
871 * track the current fast-clear and compression state. This leaves us
872 * with just MCS and CCS_E.
874 if (aux_usage
!= ISL_AUX_USAGE_CCS_E
&&
875 aux_usage
!= ISL_AUX_USAGE_MCS
)
878 set_image_compressed_bit(cmd_buffer
, image
, aspect
,
879 level
, base_layer
, layer_count
, true);
883 init_fast_clear_color(struct anv_cmd_buffer
*cmd_buffer
,
884 const struct anv_image
*image
,
885 VkImageAspectFlagBits aspect
)
887 assert(cmd_buffer
&& image
);
888 assert(image
->aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
);
890 set_image_fast_clear_state(cmd_buffer
, image
, aspect
,
891 ANV_FAST_CLEAR_NONE
);
893 /* Initialize the struct fields that are accessed for fast-clears so that
894 * the HW restrictions on the field values are satisfied.
896 struct anv_address addr
=
897 anv_image_get_clear_color_addr(cmd_buffer
->device
, image
, aspect
);
900 const struct isl_device
*isl_dev
= &cmd_buffer
->device
->isl_dev
;
901 const unsigned num_dwords
= GEN_GEN
>= 10 ?
902 isl_dev
->ss
.clear_color_state_size
/ 4 :
903 isl_dev
->ss
.clear_value_size
/ 4;
904 for (unsigned i
= 0; i
< num_dwords
; i
++) {
905 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_STORE_DATA_IMM
), sdi
) {
907 sdi
.Address
.offset
+= i
* 4;
908 sdi
.ImmediateData
= 0;
912 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_STORE_DATA_IMM
), sdi
) {
914 if (GEN_GEN
>= 8 || GEN_IS_HASWELL
) {
915 /* Pre-SKL, the dword containing the clear values also contains
916 * other fields, so we need to initialize those fields to match the
917 * values that would be in a color attachment.
919 sdi
.ImmediateData
= ISL_CHANNEL_SELECT_RED
<< 25 |
920 ISL_CHANNEL_SELECT_GREEN
<< 22 |
921 ISL_CHANNEL_SELECT_BLUE
<< 19 |
922 ISL_CHANNEL_SELECT_ALPHA
<< 16;
923 } else if (GEN_GEN
== 7) {
924 /* On IVB, the dword containing the clear values also contains
925 * other fields that must be zero or can be zero.
927 sdi
.ImmediateData
= 0;
933 /* Copy the fast-clear value dword(s) between a surface state object and an
934 * image's fast clear state buffer.
937 genX(copy_fast_clear_dwords
)(struct anv_cmd_buffer
*cmd_buffer
,
938 struct anv_state surface_state
,
939 const struct anv_image
*image
,
940 VkImageAspectFlagBits aspect
,
941 bool copy_from_surface_state
)
943 assert(cmd_buffer
&& image
);
944 assert(image
->aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
);
946 struct anv_address ss_clear_addr
= {
947 .bo
= cmd_buffer
->device
->surface_state_pool
.block_pool
.bo
,
948 .offset
= surface_state
.offset
+
949 cmd_buffer
->device
->isl_dev
.ss
.clear_value_offset
,
951 const struct anv_address entry_addr
=
952 anv_image_get_clear_color_addr(cmd_buffer
->device
, image
, aspect
);
953 unsigned copy_size
= cmd_buffer
->device
->isl_dev
.ss
.clear_value_size
;
956 /* On gen7, the combination of commands used here(MI_LOAD_REGISTER_MEM
957 * and MI_STORE_REGISTER_MEM) can cause GPU hangs if any rendering is
958 * in-flight when they are issued even if the memory touched is not
959 * currently active for rendering. The weird bit is that it is not the
960 * MI_LOAD/STORE_REGISTER_MEM commands which hang but rather the in-flight
961 * rendering hangs such that the next stalling command after the
962 * MI_LOAD/STORE_REGISTER_MEM commands will catch the hang.
964 * It is unclear exactly why this hang occurs. Both MI commands come with
965 * warnings about the 3D pipeline but that doesn't seem to fully explain
966 * it. My (Jason's) best theory is that it has something to do with the
967 * fact that we're using a GPU state register as our temporary and that
968 * something with reading/writing it is causing problems.
970 * In order to work around this issue, we emit a PIPE_CONTROL with the
971 * command streamer stall bit set.
973 cmd_buffer
->state
.pending_pipe_bits
|= ANV_PIPE_CS_STALL_BIT
;
974 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
977 struct gen_mi_builder b
;
978 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
980 if (copy_from_surface_state
) {
981 gen_mi_memcpy(&b
, entry_addr
, ss_clear_addr
, copy_size
);
983 gen_mi_memcpy(&b
, ss_clear_addr
, entry_addr
, copy_size
);
985 /* Updating a surface state object may require that the state cache be
986 * invalidated. From the SKL PRM, Shared Functions -> State -> State
989 * Whenever the RENDER_SURFACE_STATE object in memory pointed to by
990 * the Binding Table Pointer (BTP) and Binding Table Index (BTI) is
991 * modified [...], the L1 state cache must be invalidated to ensure
992 * the new surface or sampler state is fetched from system memory.
994 * In testing, SKL doesn't actually seem to need this, but HSW does.
996 cmd_buffer
->state
.pending_pipe_bits
|=
997 ANV_PIPE_STATE_CACHE_INVALIDATE_BIT
;
1001 #define READ_ONCE(x) (*(volatile __typeof__(x) *)&(x))
1005 anv_image_init_aux_tt(struct anv_cmd_buffer
*cmd_buffer
,
1006 const struct anv_image
*image
,
1007 VkImageAspectFlagBits aspect
,
1008 uint32_t base_level
, uint32_t level_count
,
1009 uint32_t base_layer
, uint32_t layer_count
)
1011 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
, aspect
);
1013 uint64_t base_address
=
1014 anv_address_physical(image
->planes
[plane
].address
);
1016 const struct isl_surf
*isl_surf
= &image
->planes
[plane
].surface
.isl
;
1017 uint64_t format_bits
= gen_aux_map_format_bits_for_isl_surf(isl_surf
);
1019 /* We're about to live-update the AUX-TT. We really don't want anyone else
1020 * trying to read it while we're doing this. We could probably get away
1021 * with not having this stall in some cases if we were really careful but
1022 * it's better to play it safe. Full stall the GPU.
1024 cmd_buffer
->state
.pending_pipe_bits
|= ANV_PIPE_CS_STALL_BIT
;
1025 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
1027 struct gen_mi_builder b
;
1028 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
1030 for (uint32_t a
= 0; a
< layer_count
; a
++) {
1031 const uint32_t layer
= base_layer
+ a
;
1033 uint64_t start_offset_B
= UINT64_MAX
, end_offset_B
= 0;
1034 for (uint32_t l
= 0; l
< level_count
; l
++) {
1035 const uint32_t level
= base_level
+ l
;
1037 uint32_t logical_array_layer
, logical_z_offset_px
;
1038 if (image
->type
== VK_IMAGE_TYPE_3D
) {
1039 logical_array_layer
= 0;
1041 /* If the given miplevel does not have this layer, then any higher
1042 * miplevels won't either because miplevels only get smaller the
1045 assert(layer
< image
->extent
.depth
);
1046 if (layer
>= anv_minify(image
->extent
.depth
, level
))
1048 logical_z_offset_px
= layer
;
1050 assert(layer
< image
->array_size
);
1051 logical_array_layer
= layer
;
1052 logical_z_offset_px
= 0;
1055 uint32_t slice_start_offset_B
, slice_end_offset_B
;
1056 isl_surf_get_image_range_B_tile(isl_surf
, level
,
1057 logical_array_layer
,
1058 logical_z_offset_px
,
1059 &slice_start_offset_B
,
1060 &slice_end_offset_B
);
1062 start_offset_B
= MIN2(start_offset_B
, slice_start_offset_B
);
1063 end_offset_B
= MAX2(end_offset_B
, slice_end_offset_B
);
1066 /* Aux operates 64K at a time */
1067 start_offset_B
= align_down_u64(start_offset_B
, 64 * 1024);
1068 end_offset_B
= align_u64(end_offset_B
, 64 * 1024);
1070 for (uint64_t offset
= start_offset_B
;
1071 offset
< end_offset_B
; offset
+= 64 * 1024) {
1072 uint64_t address
= base_address
+ offset
;
1074 uint64_t aux_entry_addr64
, *aux_entry_map
;
1075 aux_entry_map
= gen_aux_map_get_entry(cmd_buffer
->device
->aux_map_ctx
,
1076 address
, &aux_entry_addr64
);
1078 assert(cmd_buffer
->device
->physical
->use_softpin
);
1079 struct anv_address aux_entry_address
= {
1081 .offset
= aux_entry_addr64
,
1084 const uint64_t old_aux_entry
= READ_ONCE(*aux_entry_map
);
1085 uint64_t new_aux_entry
=
1086 (old_aux_entry
& GEN_AUX_MAP_ADDRESS_MASK
) | format_bits
;
1088 if (isl_aux_usage_has_ccs(image
->planes
[plane
].aux_usage
))
1089 new_aux_entry
|= GEN_AUX_MAP_ENTRY_VALID_BIT
;
1091 gen_mi_store(&b
, gen_mi_mem64(aux_entry_address
),
1092 gen_mi_imm(new_aux_entry
));
1096 cmd_buffer
->state
.pending_pipe_bits
|= ANV_PIPE_AUX_TABLE_INVALIDATE_BIT
;
1098 #endif /* GEN_GEN == 12 */
1101 * @brief Transitions a color buffer from one layout to another.
1103 * See section 6.1.1. Image Layout Transitions of the Vulkan 1.0.50 spec for
1106 * @param level_count VK_REMAINING_MIP_LEVELS isn't supported.
1107 * @param layer_count VK_REMAINING_ARRAY_LAYERS isn't supported. For 3D images,
1108 * this represents the maximum layers to transition at each
1109 * specified miplevel.
1112 transition_color_buffer(struct anv_cmd_buffer
*cmd_buffer
,
1113 const struct anv_image
*image
,
1114 VkImageAspectFlagBits aspect
,
1115 const uint32_t base_level
, uint32_t level_count
,
1116 uint32_t base_layer
, uint32_t layer_count
,
1117 VkImageLayout initial_layout
,
1118 VkImageLayout final_layout
)
1120 struct anv_device
*device
= cmd_buffer
->device
;
1121 const struct gen_device_info
*devinfo
= &device
->info
;
1122 /* Validate the inputs. */
1124 assert(image
&& image
->aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
);
1125 /* These values aren't supported for simplicity's sake. */
1126 assert(level_count
!= VK_REMAINING_MIP_LEVELS
&&
1127 layer_count
!= VK_REMAINING_ARRAY_LAYERS
);
1128 /* Ensure the subresource range is valid. */
1129 UNUSED
uint64_t last_level_num
= base_level
+ level_count
;
1130 const uint32_t max_depth
= anv_minify(image
->extent
.depth
, base_level
);
1131 UNUSED
const uint32_t image_layers
= MAX2(image
->array_size
, max_depth
);
1132 assert((uint64_t)base_layer
+ layer_count
<= image_layers
);
1133 assert(last_level_num
<= image
->levels
);
1134 /* The spec disallows these final layouts. */
1135 assert(final_layout
!= VK_IMAGE_LAYOUT_UNDEFINED
&&
1136 final_layout
!= VK_IMAGE_LAYOUT_PREINITIALIZED
);
1138 /* No work is necessary if the layout stays the same or if this subresource
1139 * range lacks auxiliary data.
1141 if (initial_layout
== final_layout
)
1144 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
, aspect
);
1146 if (image
->planes
[plane
].shadow_surface
.isl
.size_B
> 0 &&
1147 final_layout
== VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL
) {
1148 /* This surface is a linear compressed image with a tiled shadow surface
1149 * for texturing. The client is about to use it in READ_ONLY_OPTIMAL so
1150 * we need to ensure the shadow copy is up-to-date.
1152 assert(image
->aspects
== VK_IMAGE_ASPECT_COLOR_BIT
);
1153 assert(image
->planes
[plane
].surface
.isl
.tiling
== ISL_TILING_LINEAR
);
1154 assert(image
->planes
[plane
].shadow_surface
.isl
.tiling
!= ISL_TILING_LINEAR
);
1155 assert(isl_format_is_compressed(image
->planes
[plane
].surface
.isl
.format
));
1157 anv_image_copy_to_shadow(cmd_buffer
, image
,
1158 VK_IMAGE_ASPECT_COLOR_BIT
,
1159 base_level
, level_count
,
1160 base_layer
, layer_count
);
1163 if (base_layer
>= anv_image_aux_layers(image
, aspect
, base_level
))
1166 assert(image
->planes
[plane
].surface
.isl
.tiling
!= ISL_TILING_LINEAR
);
1168 if (initial_layout
== VK_IMAGE_LAYOUT_UNDEFINED
||
1169 initial_layout
== VK_IMAGE_LAYOUT_PREINITIALIZED
) {
1171 if (device
->physical
->has_implicit_ccs
&& devinfo
->has_aux_map
) {
1172 anv_image_init_aux_tt(cmd_buffer
, image
, aspect
,
1173 base_level
, level_count
,
1174 base_layer
, layer_count
);
1177 assert(!(device
->physical
->has_implicit_ccs
&& devinfo
->has_aux_map
));
1180 /* A subresource in the undefined layout may have been aliased and
1181 * populated with any arrangement of bits. Therefore, we must initialize
1182 * the related aux buffer and clear buffer entry with desirable values.
1183 * An initial layout of PREINITIALIZED is the same as UNDEFINED for
1184 * images with VK_IMAGE_TILING_OPTIMAL.
1186 * Initialize the relevant clear buffer entries.
1188 if (base_level
== 0 && base_layer
== 0)
1189 init_fast_clear_color(cmd_buffer
, image
, aspect
);
1191 /* Initialize the aux buffers to enable correct rendering. In order to
1192 * ensure that things such as storage images work correctly, aux buffers
1193 * need to be initialized to valid data.
1195 * Having an aux buffer with invalid data is a problem for two reasons:
1197 * 1) Having an invalid value in the buffer can confuse the hardware.
1198 * For instance, with CCS_E on SKL, a two-bit CCS value of 2 is
1199 * invalid and leads to the hardware doing strange things. It
1200 * doesn't hang as far as we can tell but rendering corruption can
1203 * 2) If this transition is into the GENERAL layout and we then use the
1204 * image as a storage image, then we must have the aux buffer in the
1205 * pass-through state so that, if we then go to texture from the
1206 * image, we get the results of our storage image writes and not the
1207 * fast clear color or other random data.
1209 * For CCS both of the problems above are real demonstrable issues. In
1210 * that case, the only thing we can do is to perform an ambiguate to
1211 * transition the aux surface into the pass-through state.
1213 * For MCS, (2) is never an issue because we don't support multisampled
1214 * storage images. In theory, issue (1) is a problem with MCS but we've
1215 * never seen it in the wild. For 4x and 16x, all bit patters could, in
1216 * theory, be interpreted as something but we don't know that all bit
1217 * patterns are actually valid. For 2x and 8x, you could easily end up
1218 * with the MCS referring to an invalid plane because not all bits of
1219 * the MCS value are actually used. Even though we've never seen issues
1220 * in the wild, it's best to play it safe and initialize the MCS. We
1221 * can use a fast-clear for MCS because we only ever touch from render
1222 * and texture (no image load store).
1224 if (image
->samples
== 1) {
1225 for (uint32_t l
= 0; l
< level_count
; l
++) {
1226 const uint32_t level
= base_level
+ l
;
1228 uint32_t aux_layers
= anv_image_aux_layers(image
, aspect
, level
);
1229 if (base_layer
>= aux_layers
)
1230 break; /* We will only get fewer layers as level increases */
1231 uint32_t level_layer_count
=
1232 MIN2(layer_count
, aux_layers
- base_layer
);
1234 anv_image_ccs_op(cmd_buffer
, image
,
1235 image
->planes
[plane
].surface
.isl
.format
,
1236 aspect
, level
, base_layer
, level_layer_count
,
1237 ISL_AUX_OP_AMBIGUATE
, NULL
, false);
1239 if (image
->planes
[plane
].aux_usage
== ISL_AUX_USAGE_CCS_E
) {
1240 set_image_compressed_bit(cmd_buffer
, image
, aspect
,
1241 level
, base_layer
, level_layer_count
,
1246 if (image
->samples
== 4 || image
->samples
== 16) {
1247 anv_perf_warn(cmd_buffer
->device
, image
,
1248 "Doing a potentially unnecessary fast-clear to "
1249 "define an MCS buffer.");
1252 assert(base_level
== 0 && level_count
== 1);
1253 anv_image_mcs_op(cmd_buffer
, image
,
1254 image
->planes
[plane
].surface
.isl
.format
,
1255 aspect
, base_layer
, layer_count
,
1256 ISL_AUX_OP_FAST_CLEAR
, NULL
, false);
1261 const enum isl_aux_usage initial_aux_usage
=
1262 anv_layout_to_aux_usage(devinfo
, image
, aspect
, 0, initial_layout
);
1263 const enum isl_aux_usage final_aux_usage
=
1264 anv_layout_to_aux_usage(devinfo
, image
, aspect
, 0, final_layout
);
1266 /* The current code assumes that there is no mixing of CCS_E and CCS_D.
1267 * We can handle transitions between CCS_D/E to and from NONE. What we
1268 * don't yet handle is switching between CCS_E and CCS_D within a given
1269 * image. Doing so in a performant way requires more detailed aux state
1270 * tracking such as what is done in i965. For now, just assume that we
1271 * only have one type of compression.
1273 assert(initial_aux_usage
== ISL_AUX_USAGE_NONE
||
1274 final_aux_usage
== ISL_AUX_USAGE_NONE
||
1275 initial_aux_usage
== final_aux_usage
);
1277 /* If initial aux usage is NONE, there is nothing to resolve */
1278 if (initial_aux_usage
== ISL_AUX_USAGE_NONE
)
1281 enum isl_aux_op resolve_op
= ISL_AUX_OP_NONE
;
1283 /* If the initial layout supports more fast clear than the final layout
1284 * then we need at least a partial resolve.
1286 const enum anv_fast_clear_type initial_fast_clear
=
1287 anv_layout_to_fast_clear_type(devinfo
, image
, aspect
, initial_layout
);
1288 const enum anv_fast_clear_type final_fast_clear
=
1289 anv_layout_to_fast_clear_type(devinfo
, image
, aspect
, final_layout
);
1290 if (final_fast_clear
< initial_fast_clear
)
1291 resolve_op
= ISL_AUX_OP_PARTIAL_RESOLVE
;
1293 if (initial_aux_usage
== ISL_AUX_USAGE_CCS_E
&&
1294 final_aux_usage
!= ISL_AUX_USAGE_CCS_E
)
1295 resolve_op
= ISL_AUX_OP_FULL_RESOLVE
;
1297 if (resolve_op
== ISL_AUX_OP_NONE
)
1300 /* Perform a resolve to synchronize data between the main and aux buffer.
1301 * Before we begin, we must satisfy the cache flushing requirement specified
1302 * in the Sky Lake PRM Vol. 7, "MCS Buffer for Render Target(s)":
1304 * Any transition from any value in {Clear, Render, Resolve} to a
1305 * different value in {Clear, Render, Resolve} requires end of pipe
1308 * We perform a flush of the write cache before and after the clear and
1309 * resolve operations to meet this requirement.
1311 * Unlike other drawing, fast clear operations are not properly
1312 * synchronized. The first PIPE_CONTROL here likely ensures that the
1313 * contents of the previous render or clear hit the render target before we
1314 * resolve and the second likely ensures that the resolve is complete before
1315 * we do any more rendering or clearing.
1317 cmd_buffer
->state
.pending_pipe_bits
|=
1318 ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT
| ANV_PIPE_END_OF_PIPE_SYNC_BIT
;
1320 for (uint32_t l
= 0; l
< level_count
; l
++) {
1321 uint32_t level
= base_level
+ l
;
1323 uint32_t aux_layers
= anv_image_aux_layers(image
, aspect
, level
);
1324 if (base_layer
>= aux_layers
)
1325 break; /* We will only get fewer layers as level increases */
1326 uint32_t level_layer_count
=
1327 MIN2(layer_count
, aux_layers
- base_layer
);
1329 for (uint32_t a
= 0; a
< level_layer_count
; a
++) {
1330 uint32_t array_layer
= base_layer
+ a
;
1331 if (image
->samples
== 1) {
1332 anv_cmd_predicated_ccs_resolve(cmd_buffer
, image
,
1333 image
->planes
[plane
].surface
.isl
.format
,
1334 aspect
, level
, array_layer
, resolve_op
,
1337 /* We only support fast-clear on the first layer so partial
1338 * resolves should not be used on other layers as they will use
1339 * the clear color stored in memory that is only valid for layer0.
1341 if (resolve_op
== ISL_AUX_OP_PARTIAL_RESOLVE
&&
1345 anv_cmd_predicated_mcs_resolve(cmd_buffer
, image
,
1346 image
->planes
[plane
].surface
.isl
.format
,
1347 aspect
, array_layer
, resolve_op
,
1353 cmd_buffer
->state
.pending_pipe_bits
|=
1354 ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT
| ANV_PIPE_END_OF_PIPE_SYNC_BIT
;
1358 * Setup anv_cmd_state::attachments for vkCmdBeginRenderPass.
1361 genX(cmd_buffer_setup_attachments
)(struct anv_cmd_buffer
*cmd_buffer
,
1362 struct anv_render_pass
*pass
,
1363 const VkRenderPassBeginInfo
*begin
)
1365 const struct isl_device
*isl_dev
= &cmd_buffer
->device
->isl_dev
;
1366 struct anv_cmd_state
*state
= &cmd_buffer
->state
;
1367 struct anv_framebuffer
*framebuffer
= cmd_buffer
->state
.framebuffer
;
1369 vk_free(&cmd_buffer
->pool
->alloc
, state
->attachments
);
1371 if (pass
->attachment_count
> 0) {
1372 state
->attachments
= vk_alloc(&cmd_buffer
->pool
->alloc
,
1373 pass
->attachment_count
*
1374 sizeof(state
->attachments
[0]),
1375 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1376 if (state
->attachments
== NULL
) {
1377 /* Propagate VK_ERROR_OUT_OF_HOST_MEMORY to vkEndCommandBuffer */
1378 return anv_batch_set_error(&cmd_buffer
->batch
,
1379 VK_ERROR_OUT_OF_HOST_MEMORY
);
1382 state
->attachments
= NULL
;
1385 /* Reserve one for the NULL state. */
1386 unsigned num_states
= 1;
1387 for (uint32_t i
= 0; i
< pass
->attachment_count
; ++i
) {
1388 if (vk_format_is_color(pass
->attachments
[i
].format
))
1391 if (need_input_attachment_state(&pass
->attachments
[i
]))
1395 const uint32_t ss_stride
= align_u32(isl_dev
->ss
.size
, isl_dev
->ss
.align
);
1396 state
->render_pass_states
=
1397 anv_state_stream_alloc(&cmd_buffer
->surface_state_stream
,
1398 num_states
* ss_stride
, isl_dev
->ss
.align
);
1400 struct anv_state next_state
= state
->render_pass_states
;
1401 next_state
.alloc_size
= isl_dev
->ss
.size
;
1403 state
->null_surface_state
= next_state
;
1404 next_state
.offset
+= ss_stride
;
1405 next_state
.map
+= ss_stride
;
1407 const VkRenderPassAttachmentBeginInfoKHR
*begin_attachment
=
1408 vk_find_struct_const(begin
, RENDER_PASS_ATTACHMENT_BEGIN_INFO_KHR
);
1410 if (begin
&& !begin_attachment
)
1411 assert(pass
->attachment_count
== framebuffer
->attachment_count
);
1413 for (uint32_t i
= 0; i
< pass
->attachment_count
; ++i
) {
1414 if (vk_format_is_color(pass
->attachments
[i
].format
)) {
1415 state
->attachments
[i
].color
.state
= next_state
;
1416 next_state
.offset
+= ss_stride
;
1417 next_state
.map
+= ss_stride
;
1420 if (need_input_attachment_state(&pass
->attachments
[i
])) {
1421 state
->attachments
[i
].input
.state
= next_state
;
1422 next_state
.offset
+= ss_stride
;
1423 next_state
.map
+= ss_stride
;
1426 if (begin_attachment
&& begin_attachment
->attachmentCount
!= 0) {
1427 assert(begin_attachment
->attachmentCount
== pass
->attachment_count
);
1428 ANV_FROM_HANDLE(anv_image_view
, iview
, begin_attachment
->pAttachments
[i
]);
1429 cmd_buffer
->state
.attachments
[i
].image_view
= iview
;
1430 } else if (framebuffer
&& i
< framebuffer
->attachment_count
) {
1431 cmd_buffer
->state
.attachments
[i
].image_view
= framebuffer
->attachments
[i
];
1434 assert(next_state
.offset
== state
->render_pass_states
.offset
+
1435 state
->render_pass_states
.alloc_size
);
1438 isl_null_fill_state(isl_dev
, state
->null_surface_state
.map
,
1439 isl_extent3d(framebuffer
->width
,
1440 framebuffer
->height
,
1441 framebuffer
->layers
));
1443 for (uint32_t i
= 0; i
< pass
->attachment_count
; ++i
) {
1444 struct anv_render_pass_attachment
*att
= &pass
->attachments
[i
];
1445 VkImageAspectFlags att_aspects
= vk_format_aspects(att
->format
);
1446 VkImageAspectFlags clear_aspects
= 0;
1447 VkImageAspectFlags load_aspects
= 0;
1449 if (att_aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) {
1450 /* color attachment */
1451 if (att
->load_op
== VK_ATTACHMENT_LOAD_OP_CLEAR
) {
1452 clear_aspects
|= VK_IMAGE_ASPECT_COLOR_BIT
;
1453 } else if (att
->load_op
== VK_ATTACHMENT_LOAD_OP_LOAD
) {
1454 load_aspects
|= VK_IMAGE_ASPECT_COLOR_BIT
;
1457 /* depthstencil attachment */
1458 if (att_aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
) {
1459 if (att
->load_op
== VK_ATTACHMENT_LOAD_OP_CLEAR
) {
1460 clear_aspects
|= VK_IMAGE_ASPECT_DEPTH_BIT
;
1461 } else if (att
->load_op
== VK_ATTACHMENT_LOAD_OP_LOAD
) {
1462 load_aspects
|= VK_IMAGE_ASPECT_DEPTH_BIT
;
1465 if (att_aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
) {
1466 if (att
->stencil_load_op
== VK_ATTACHMENT_LOAD_OP_CLEAR
) {
1467 clear_aspects
|= VK_IMAGE_ASPECT_STENCIL_BIT
;
1468 } else if (att
->stencil_load_op
== VK_ATTACHMENT_LOAD_OP_LOAD
) {
1469 load_aspects
|= VK_IMAGE_ASPECT_STENCIL_BIT
;
1474 state
->attachments
[i
].current_layout
= att
->initial_layout
;
1475 state
->attachments
[i
].current_stencil_layout
= att
->stencil_initial_layout
;
1476 state
->attachments
[i
].pending_clear_aspects
= clear_aspects
;
1477 state
->attachments
[i
].pending_load_aspects
= load_aspects
;
1479 state
->attachments
[i
].clear_value
= begin
->pClearValues
[i
];
1481 struct anv_image_view
*iview
= cmd_buffer
->state
.attachments
[i
].image_view
;
1482 anv_assert(iview
->vk_format
== att
->format
);
1484 const uint32_t num_layers
= iview
->planes
[0].isl
.array_len
;
1485 state
->attachments
[i
].pending_clear_views
= (1 << num_layers
) - 1;
1487 union isl_color_value clear_color
= { .u32
= { 0, } };
1488 if (att_aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) {
1489 anv_assert(iview
->n_planes
== 1);
1490 assert(att_aspects
== VK_IMAGE_ASPECT_COLOR_BIT
);
1491 color_attachment_compute_aux_usage(cmd_buffer
->device
,
1492 state
, i
, begin
->renderArea
,
1495 anv_image_fill_surface_state(cmd_buffer
->device
,
1497 VK_IMAGE_ASPECT_COLOR_BIT
,
1498 &iview
->planes
[0].isl
,
1499 ISL_SURF_USAGE_RENDER_TARGET_BIT
,
1500 state
->attachments
[i
].aux_usage
,
1503 &state
->attachments
[i
].color
,
1506 add_surface_state_relocs(cmd_buffer
, state
->attachments
[i
].color
);
1508 depth_stencil_attachment_compute_aux_usage(cmd_buffer
->device
,
1513 if (need_input_attachment_state(&pass
->attachments
[i
])) {
1514 anv_image_fill_surface_state(cmd_buffer
->device
,
1516 VK_IMAGE_ASPECT_COLOR_BIT
,
1517 &iview
->planes
[0].isl
,
1518 ISL_SURF_USAGE_TEXTURE_BIT
,
1519 state
->attachments
[i
].input_aux_usage
,
1522 &state
->attachments
[i
].input
,
1525 add_surface_state_relocs(cmd_buffer
, state
->attachments
[i
].input
);
1534 genX(BeginCommandBuffer
)(
1535 VkCommandBuffer commandBuffer
,
1536 const VkCommandBufferBeginInfo
* pBeginInfo
)
1538 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
1540 /* If this is the first vkBeginCommandBuffer, we must *initialize* the
1541 * command buffer's state. Otherwise, we must *reset* its state. In both
1542 * cases we reset it.
1544 * From the Vulkan 1.0 spec:
1546 * If a command buffer is in the executable state and the command buffer
1547 * was allocated from a command pool with the
1548 * VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT flag set, then
1549 * vkBeginCommandBuffer implicitly resets the command buffer, behaving
1550 * as if vkResetCommandBuffer had been called with
1551 * VK_COMMAND_BUFFER_RESET_RELEASE_RESOURCES_BIT not set. It then puts
1552 * the command buffer in the recording state.
1554 anv_cmd_buffer_reset(cmd_buffer
);
1556 cmd_buffer
->usage_flags
= pBeginInfo
->flags
;
1558 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_SECONDARY
||
1559 !(cmd_buffer
->usage_flags
& VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT
));
1561 genX(cmd_buffer_emit_state_base_address
)(cmd_buffer
);
1563 /* We sometimes store vertex data in the dynamic state buffer for blorp
1564 * operations and our dynamic state stream may re-use data from previous
1565 * command buffers. In order to prevent stale cache data, we flush the VF
1566 * cache. We could do this on every blorp call but that's not really
1567 * needed as all of the data will get written by the CPU prior to the GPU
1568 * executing anything. The chances are fairly high that they will use
1569 * blorp at least once per primary command buffer so it shouldn't be
1572 * There is also a workaround on gen8 which requires us to invalidate the
1573 * VF cache occasionally. It's easier if we can assume we start with a
1574 * fresh cache (See also genX(cmd_buffer_set_binding_for_gen8_vb_flush).)
1576 cmd_buffer
->state
.pending_pipe_bits
|= ANV_PIPE_VF_CACHE_INVALIDATE_BIT
;
1578 /* Re-emit the aux table register in every command buffer. This way we're
1579 * ensured that we have the table even if this command buffer doesn't
1580 * initialize any images.
1582 cmd_buffer
->state
.pending_pipe_bits
|= ANV_PIPE_AUX_TABLE_INVALIDATE_BIT
;
1584 /* We send an "Indirect State Pointers Disable" packet at
1585 * EndCommandBuffer, so all push contant packets are ignored during a
1586 * context restore. Documentation says after that command, we need to
1587 * emit push constants again before any rendering operation. So we
1588 * flag them dirty here to make sure they get emitted.
1590 cmd_buffer
->state
.push_constants_dirty
|= VK_SHADER_STAGE_ALL_GRAPHICS
;
1592 VkResult result
= VK_SUCCESS
;
1593 if (cmd_buffer
->usage_flags
&
1594 VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT
) {
1595 assert(pBeginInfo
->pInheritanceInfo
);
1596 cmd_buffer
->state
.pass
=
1597 anv_render_pass_from_handle(pBeginInfo
->pInheritanceInfo
->renderPass
);
1598 cmd_buffer
->state
.subpass
=
1599 &cmd_buffer
->state
.pass
->subpasses
[pBeginInfo
->pInheritanceInfo
->subpass
];
1601 /* This is optional in the inheritance info. */
1602 cmd_buffer
->state
.framebuffer
=
1603 anv_framebuffer_from_handle(pBeginInfo
->pInheritanceInfo
->framebuffer
);
1605 result
= genX(cmd_buffer_setup_attachments
)(cmd_buffer
,
1606 cmd_buffer
->state
.pass
, NULL
);
1608 /* Record that HiZ is enabled if we can. */
1609 if (cmd_buffer
->state
.framebuffer
) {
1610 const struct anv_image_view
* const iview
=
1611 anv_cmd_buffer_get_depth_stencil_view(cmd_buffer
);
1614 VkImageLayout layout
=
1615 cmd_buffer
->state
.subpass
->depth_stencil_attachment
->layout
;
1617 enum isl_aux_usage aux_usage
=
1618 anv_layout_to_aux_usage(&cmd_buffer
->device
->info
, iview
->image
,
1619 VK_IMAGE_ASPECT_DEPTH_BIT
,
1620 VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT
,
1623 cmd_buffer
->state
.hiz_enabled
= aux_usage
== ISL_AUX_USAGE_HIZ
;
1627 cmd_buffer
->state
.gfx
.dirty
|= ANV_CMD_DIRTY_RENDER_TARGETS
;
1630 #if GEN_GEN >= 8 || GEN_IS_HASWELL
1631 if (cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_SECONDARY
) {
1632 const VkCommandBufferInheritanceConditionalRenderingInfoEXT
*conditional_rendering_info
=
1633 vk_find_struct_const(pBeginInfo
->pInheritanceInfo
->pNext
, COMMAND_BUFFER_INHERITANCE_CONDITIONAL_RENDERING_INFO_EXT
);
1635 /* If secondary buffer supports conditional rendering
1636 * we should emit commands as if conditional rendering is enabled.
1638 cmd_buffer
->state
.conditional_render_enabled
=
1639 conditional_rendering_info
&& conditional_rendering_info
->conditionalRenderingEnable
;
1646 /* From the PRM, Volume 2a:
1648 * "Indirect State Pointers Disable
1650 * At the completion of the post-sync operation associated with this pipe
1651 * control packet, the indirect state pointers in the hardware are
1652 * considered invalid; the indirect pointers are not saved in the context.
1653 * If any new indirect state commands are executed in the command stream
1654 * while the pipe control is pending, the new indirect state commands are
1657 * [DevIVB+]: Using Invalidate State Pointer (ISP) only inhibits context
1658 * restoring of Push Constant (3DSTATE_CONSTANT_*) commands. Push Constant
1659 * commands are only considered as Indirect State Pointers. Once ISP is
1660 * issued in a context, SW must initialize by programming push constant
1661 * commands for all the shaders (at least to zero length) before attempting
1662 * any rendering operation for the same context."
1664 * 3DSTATE_CONSTANT_* packets are restored during a context restore,
1665 * even though they point to a BO that has been already unreferenced at
1666 * the end of the previous batch buffer. This has been fine so far since
1667 * we are protected by these scratch page (every address not covered by
1668 * a BO should be pointing to the scratch page). But on CNL, it is
1669 * causing a GPU hang during context restore at the 3DSTATE_CONSTANT_*
1672 * The flag "Indirect State Pointers Disable" in PIPE_CONTROL tells the
1673 * hardware to ignore previous 3DSTATE_CONSTANT_* packets during a
1674 * context restore, so the mentioned hang doesn't happen. However,
1675 * software must program push constant commands for all stages prior to
1676 * rendering anything. So we flag them dirty in BeginCommandBuffer.
1678 * Finally, we also make sure to stall at pixel scoreboard to make sure the
1679 * constants have been loaded into the EUs prior to disable the push constants
1680 * so that it doesn't hang a previous 3DPRIMITIVE.
1683 emit_isp_disable(struct anv_cmd_buffer
*cmd_buffer
)
1685 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
1686 pc
.StallAtPixelScoreboard
= true;
1687 pc
.CommandStreamerStallEnable
= true;
1689 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
1690 pc
.IndirectStatePointersDisable
= true;
1691 pc
.CommandStreamerStallEnable
= true;
1696 genX(EndCommandBuffer
)(
1697 VkCommandBuffer commandBuffer
)
1699 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
1701 if (anv_batch_has_error(&cmd_buffer
->batch
))
1702 return cmd_buffer
->batch
.status
;
1704 /* We want every command buffer to start with the PMA fix in a known state,
1705 * so we disable it at the end of the command buffer.
1707 genX(cmd_buffer_enable_pma_fix
)(cmd_buffer
, false);
1709 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
1711 emit_isp_disable(cmd_buffer
);
1713 anv_cmd_buffer_end_batch_buffer(cmd_buffer
);
1719 genX(CmdExecuteCommands
)(
1720 VkCommandBuffer commandBuffer
,
1721 uint32_t commandBufferCount
,
1722 const VkCommandBuffer
* pCmdBuffers
)
1724 ANV_FROM_HANDLE(anv_cmd_buffer
, primary
, commandBuffer
);
1726 assert(primary
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
1728 if (anv_batch_has_error(&primary
->batch
))
1731 /* The secondary command buffers will assume that the PMA fix is disabled
1732 * when they begin executing. Make sure this is true.
1734 genX(cmd_buffer_enable_pma_fix
)(primary
, false);
1736 /* The secondary command buffer doesn't know which textures etc. have been
1737 * flushed prior to their execution. Apply those flushes now.
1739 genX(cmd_buffer_apply_pipe_flushes
)(primary
);
1741 for (uint32_t i
= 0; i
< commandBufferCount
; i
++) {
1742 ANV_FROM_HANDLE(anv_cmd_buffer
, secondary
, pCmdBuffers
[i
]);
1744 assert(secondary
->level
== VK_COMMAND_BUFFER_LEVEL_SECONDARY
);
1745 assert(!anv_batch_has_error(&secondary
->batch
));
1747 #if GEN_GEN >= 8 || GEN_IS_HASWELL
1748 if (secondary
->state
.conditional_render_enabled
) {
1749 if (!primary
->state
.conditional_render_enabled
) {
1750 /* Secondary buffer is constructed as if it will be executed
1751 * with conditional rendering, we should satisfy this dependency
1752 * regardless of conditional rendering being enabled in primary.
1754 struct gen_mi_builder b
;
1755 gen_mi_builder_init(&b
, &primary
->batch
);
1756 gen_mi_store(&b
, gen_mi_reg64(ANV_PREDICATE_RESULT_REG
),
1757 gen_mi_imm(UINT64_MAX
));
1762 if (secondary
->usage_flags
&
1763 VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT
) {
1764 /* If we're continuing a render pass from the primary, we need to
1765 * copy the surface states for the current subpass into the storage
1766 * we allocated for them in BeginCommandBuffer.
1768 struct anv_bo
*ss_bo
=
1769 primary
->device
->surface_state_pool
.block_pool
.bo
;
1770 struct anv_state src_state
= primary
->state
.render_pass_states
;
1771 struct anv_state dst_state
= secondary
->state
.render_pass_states
;
1772 assert(src_state
.alloc_size
== dst_state
.alloc_size
);
1774 genX(cmd_buffer_so_memcpy
)(primary
,
1775 (struct anv_address
) {
1777 .offset
= dst_state
.offset
,
1779 (struct anv_address
) {
1781 .offset
= src_state
.offset
,
1783 src_state
.alloc_size
);
1786 anv_cmd_buffer_add_secondary(primary
, secondary
);
1789 /* The secondary isn't counted in our VF cache tracking so we need to
1790 * invalidate the whole thing.
1792 if (GEN_GEN
>= 8 && GEN_GEN
<= 9) {
1793 primary
->state
.pending_pipe_bits
|=
1794 ANV_PIPE_CS_STALL_BIT
| ANV_PIPE_VF_CACHE_INVALIDATE_BIT
;
1797 /* The secondary may have selected a different pipeline (3D or compute) and
1798 * may have changed the current L3$ configuration. Reset our tracking
1799 * variables to invalid values to ensure that we re-emit these in the case
1800 * where we do any draws or compute dispatches from the primary after the
1801 * secondary has returned.
1803 primary
->state
.current_pipeline
= UINT32_MAX
;
1804 primary
->state
.current_l3_config
= NULL
;
1805 primary
->state
.current_hash_scale
= 0;
1807 /* Each of the secondary command buffers will use its own state base
1808 * address. We need to re-emit state base address for the primary after
1809 * all of the secondaries are done.
1811 * TODO: Maybe we want to make this a dirty bit to avoid extra state base
1814 genX(cmd_buffer_emit_state_base_address
)(primary
);
1817 #define IVB_L3SQCREG1_SQGHPCI_DEFAULT 0x00730000
1818 #define VLV_L3SQCREG1_SQGHPCI_DEFAULT 0x00d30000
1819 #define HSW_L3SQCREG1_SQGHPCI_DEFAULT 0x00610000
1822 * Program the hardware to use the specified L3 configuration.
1825 genX(cmd_buffer_config_l3
)(struct anv_cmd_buffer
*cmd_buffer
,
1826 const struct gen_l3_config
*cfg
)
1829 if (cfg
== cmd_buffer
->state
.current_l3_config
)
1832 if (unlikely(INTEL_DEBUG
& DEBUG_L3
)) {
1833 intel_logd("L3 config transition: ");
1834 gen_dump_l3_config(cfg
, stderr
);
1837 UNUSED
const bool has_slm
= cfg
->n
[GEN_L3P_SLM
];
1839 /* According to the hardware docs, the L3 partitioning can only be changed
1840 * while the pipeline is completely drained and the caches are flushed,
1841 * which involves a first PIPE_CONTROL flush which stalls the pipeline...
1843 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
1844 pc
.DCFlushEnable
= true;
1845 pc
.PostSyncOperation
= NoWrite
;
1846 pc
.CommandStreamerStallEnable
= true;
1849 /* ...followed by a second pipelined PIPE_CONTROL that initiates
1850 * invalidation of the relevant caches. Note that because RO invalidation
1851 * happens at the top of the pipeline (i.e. right away as the PIPE_CONTROL
1852 * command is processed by the CS) we cannot combine it with the previous
1853 * stalling flush as the hardware documentation suggests, because that
1854 * would cause the CS to stall on previous rendering *after* RO
1855 * invalidation and wouldn't prevent the RO caches from being polluted by
1856 * concurrent rendering before the stall completes. This intentionally
1857 * doesn't implement the SKL+ hardware workaround suggesting to enable CS
1858 * stall on PIPE_CONTROLs with the texture cache invalidation bit set for
1859 * GPGPU workloads because the previous and subsequent PIPE_CONTROLs
1860 * already guarantee that there is no concurrent GPGPU kernel execution
1861 * (see SKL HSD 2132585).
1863 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
1864 pc
.TextureCacheInvalidationEnable
= true;
1865 pc
.ConstantCacheInvalidationEnable
= true;
1866 pc
.InstructionCacheInvalidateEnable
= true;
1867 pc
.StateCacheInvalidationEnable
= true;
1868 pc
.PostSyncOperation
= NoWrite
;
1871 /* Now send a third stalling flush to make sure that invalidation is
1872 * complete when the L3 configuration registers are modified.
1874 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
1875 pc
.DCFlushEnable
= true;
1876 pc
.PostSyncOperation
= NoWrite
;
1877 pc
.CommandStreamerStallEnable
= true;
1882 assert(!cfg
->n
[GEN_L3P_IS
] && !cfg
->n
[GEN_L3P_C
] && !cfg
->n
[GEN_L3P_T
]);
1885 #define L3_ALLOCATION_REG GENX(L3ALLOC)
1886 #define L3_ALLOCATION_REG_num GENX(L3ALLOC_num)
1888 #define L3_ALLOCATION_REG GENX(L3CNTLREG)
1889 #define L3_ALLOCATION_REG_num GENX(L3CNTLREG_num)
1893 anv_pack_struct(&l3cr
, L3_ALLOCATION_REG
,
1895 .SLMEnable
= has_slm
,
1898 /* WA_1406697149: Bit 9 "Error Detection Behavior Control" must be set
1899 * in L3CNTLREG register. The default setting of the bit is not the
1900 * desirable behavior.
1902 .ErrorDetectionBehaviorControl
= true,
1903 .UseFullWays
= true,
1905 .URBAllocation
= cfg
->n
[GEN_L3P_URB
],
1906 .ROAllocation
= cfg
->n
[GEN_L3P_RO
],
1907 .DCAllocation
= cfg
->n
[GEN_L3P_DC
],
1908 .AllAllocation
= cfg
->n
[GEN_L3P_ALL
]);
1910 /* Set up the L3 partitioning. */
1911 emit_lri(&cmd_buffer
->batch
, L3_ALLOCATION_REG_num
, l3cr
);
1915 const bool has_dc
= cfg
->n
[GEN_L3P_DC
] || cfg
->n
[GEN_L3P_ALL
];
1916 const bool has_is
= cfg
->n
[GEN_L3P_IS
] || cfg
->n
[GEN_L3P_RO
] ||
1917 cfg
->n
[GEN_L3P_ALL
];
1918 const bool has_c
= cfg
->n
[GEN_L3P_C
] || cfg
->n
[GEN_L3P_RO
] ||
1919 cfg
->n
[GEN_L3P_ALL
];
1920 const bool has_t
= cfg
->n
[GEN_L3P_T
] || cfg
->n
[GEN_L3P_RO
] ||
1921 cfg
->n
[GEN_L3P_ALL
];
1923 assert(!cfg
->n
[GEN_L3P_ALL
]);
1925 /* When enabled SLM only uses a portion of the L3 on half of the banks,
1926 * the matching space on the remaining banks has to be allocated to a
1927 * client (URB for all validated configurations) set to the
1928 * lower-bandwidth 2-bank address hashing mode.
1930 const struct gen_device_info
*devinfo
= &cmd_buffer
->device
->info
;
1931 const bool urb_low_bw
= has_slm
&& !devinfo
->is_baytrail
;
1932 assert(!urb_low_bw
|| cfg
->n
[GEN_L3P_URB
] == cfg
->n
[GEN_L3P_SLM
]);
1934 /* Minimum number of ways that can be allocated to the URB. */
1935 const unsigned n0_urb
= devinfo
->is_baytrail
? 32 : 0;
1936 assert(cfg
->n
[GEN_L3P_URB
] >= n0_urb
);
1938 uint32_t l3sqcr1
, l3cr2
, l3cr3
;
1939 anv_pack_struct(&l3sqcr1
, GENX(L3SQCREG1
),
1940 .ConvertDC_UC
= !has_dc
,
1941 .ConvertIS_UC
= !has_is
,
1942 .ConvertC_UC
= !has_c
,
1943 .ConvertT_UC
= !has_t
);
1945 GEN_IS_HASWELL
? HSW_L3SQCREG1_SQGHPCI_DEFAULT
:
1946 devinfo
->is_baytrail
? VLV_L3SQCREG1_SQGHPCI_DEFAULT
:
1947 IVB_L3SQCREG1_SQGHPCI_DEFAULT
;
1949 anv_pack_struct(&l3cr2
, GENX(L3CNTLREG2
),
1950 .SLMEnable
= has_slm
,
1951 .URBLowBandwidth
= urb_low_bw
,
1952 .URBAllocation
= cfg
->n
[GEN_L3P_URB
] - n0_urb
,
1954 .ALLAllocation
= cfg
->n
[GEN_L3P_ALL
],
1956 .ROAllocation
= cfg
->n
[GEN_L3P_RO
],
1957 .DCAllocation
= cfg
->n
[GEN_L3P_DC
]);
1959 anv_pack_struct(&l3cr3
, GENX(L3CNTLREG3
),
1960 .ISAllocation
= cfg
->n
[GEN_L3P_IS
],
1961 .ISLowBandwidth
= 0,
1962 .CAllocation
= cfg
->n
[GEN_L3P_C
],
1964 .TAllocation
= cfg
->n
[GEN_L3P_T
],
1965 .TLowBandwidth
= 0);
1967 /* Set up the L3 partitioning. */
1968 emit_lri(&cmd_buffer
->batch
, GENX(L3SQCREG1_num
), l3sqcr1
);
1969 emit_lri(&cmd_buffer
->batch
, GENX(L3CNTLREG2_num
), l3cr2
);
1970 emit_lri(&cmd_buffer
->batch
, GENX(L3CNTLREG3_num
), l3cr3
);
1973 if (cmd_buffer
->device
->physical
->cmd_parser_version
>= 4) {
1974 /* Enable L3 atomics on HSW if we have a DC partition, otherwise keep
1975 * them disabled to avoid crashing the system hard.
1977 uint32_t scratch1
, chicken3
;
1978 anv_pack_struct(&scratch1
, GENX(SCRATCH1
),
1979 .L3AtomicDisable
= !has_dc
);
1980 anv_pack_struct(&chicken3
, GENX(CHICKEN3
),
1981 .L3AtomicDisableMask
= true,
1982 .L3AtomicDisable
= !has_dc
);
1983 emit_lri(&cmd_buffer
->batch
, GENX(SCRATCH1_num
), scratch1
);
1984 emit_lri(&cmd_buffer
->batch
, GENX(CHICKEN3_num
), chicken3
);
1990 cmd_buffer
->state
.current_l3_config
= cfg
;
1994 genX(cmd_buffer_apply_pipe_flushes
)(struct anv_cmd_buffer
*cmd_buffer
)
1996 UNUSED
const struct gen_device_info
*devinfo
= &cmd_buffer
->device
->info
;
1997 enum anv_pipe_bits bits
= cmd_buffer
->state
.pending_pipe_bits
;
1999 if (cmd_buffer
->device
->physical
->always_flush_cache
)
2000 bits
|= ANV_PIPE_FLUSH_BITS
| ANV_PIPE_INVALIDATE_BITS
;
2003 * From Sandybridge PRM, volume 2, "1.7.2 End-of-Pipe Synchronization":
2005 * Write synchronization is a special case of end-of-pipe
2006 * synchronization that requires that the render cache and/or depth
2007 * related caches are flushed to memory, where the data will become
2008 * globally visible. This type of synchronization is required prior to
2009 * SW (CPU) actually reading the result data from memory, or initiating
2010 * an operation that will use as a read surface (such as a texture
2011 * surface) a previous render target and/or depth/stencil buffer
2014 * From Haswell PRM, volume 2, part 1, "End-of-Pipe Synchronization":
2016 * Exercising the write cache flush bits (Render Target Cache Flush
2017 * Enable, Depth Cache Flush Enable, DC Flush) in PIPE_CONTROL only
2018 * ensures the write caches are flushed and doesn't guarantee the data
2019 * is globally visible.
2021 * SW can track the completion of the end-of-pipe-synchronization by
2022 * using "Notify Enable" and "PostSync Operation - Write Immediate
2023 * Data" in the PIPE_CONTROL command.
2025 * In other words, flushes are pipelined while invalidations are handled
2026 * immediately. Therefore, if we're flushing anything then we need to
2027 * schedule an end-of-pipe sync before any invalidations can happen.
2029 if (bits
& ANV_PIPE_FLUSH_BITS
)
2030 bits
|= ANV_PIPE_NEEDS_END_OF_PIPE_SYNC_BIT
;
2033 /* HSD 1209978178: docs say that before programming the aux table:
2035 * "Driver must ensure that the engine is IDLE but ensure it doesn't
2036 * add extra flushes in the case it knows that the engine is already
2039 if (GEN_GEN
== 12 && ANV_PIPE_AUX_TABLE_INVALIDATE_BIT
)
2040 bits
|= ANV_PIPE_NEEDS_END_OF_PIPE_SYNC_BIT
;
2042 /* If we're going to do an invalidate and we have a pending end-of-pipe
2043 * sync that has yet to be resolved, we do the end-of-pipe sync now.
2045 if ((bits
& ANV_PIPE_INVALIDATE_BITS
) &&
2046 (bits
& ANV_PIPE_NEEDS_END_OF_PIPE_SYNC_BIT
)) {
2047 bits
|= ANV_PIPE_END_OF_PIPE_SYNC_BIT
;
2048 bits
&= ~ANV_PIPE_NEEDS_END_OF_PIPE_SYNC_BIT
;
2051 if (GEN_GEN
>= 12 &&
2052 ((bits
& ANV_PIPE_DEPTH_CACHE_FLUSH_BIT
) ||
2053 (bits
& ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT
))) {
2054 /* From the PIPE_CONTROL instruction table, bit 28 (Tile Cache Flush
2057 * Unified Cache (Tile Cache Disabled):
2059 * When the Color and Depth (Z) streams are enabled to be cached in
2060 * the DC space of L2, Software must use "Render Target Cache Flush
2061 * Enable" and "Depth Cache Flush Enable" along with "Tile Cache
2062 * Flush" for getting the color and depth (Z) write data to be
2063 * globally observable. In this mode of operation it is not required
2064 * to set "CS Stall" upon setting "Tile Cache Flush" bit.
2066 bits
|= ANV_PIPE_TILE_CACHE_FLUSH_BIT
;
2069 /* GEN:BUG:1409226450, Wait for EU to be idle before pipe control which
2070 * invalidates the instruction cache
2072 if (GEN_GEN
== 12 && (bits
& ANV_PIPE_INSTRUCTION_CACHE_INVALIDATE_BIT
))
2073 bits
|= ANV_PIPE_CS_STALL_BIT
| ANV_PIPE_STALL_AT_SCOREBOARD_BIT
;
2075 if ((GEN_GEN
>= 8 && GEN_GEN
<= 9) &&
2076 (bits
& ANV_PIPE_CS_STALL_BIT
) &&
2077 (bits
& ANV_PIPE_VF_CACHE_INVALIDATE_BIT
)) {
2078 /* If we are doing a VF cache invalidate AND a CS stall (it must be
2079 * both) then we can reset our vertex cache tracking.
2081 memset(cmd_buffer
->state
.gfx
.vb_dirty_ranges
, 0,
2082 sizeof(cmd_buffer
->state
.gfx
.vb_dirty_ranges
));
2083 memset(&cmd_buffer
->state
.gfx
.ib_dirty_range
, 0,
2084 sizeof(cmd_buffer
->state
.gfx
.ib_dirty_range
));
2087 /* Project: SKL / Argument: LRI Post Sync Operation [23]
2089 * "PIPECONTROL command with “Command Streamer Stall Enable” must be
2090 * programmed prior to programming a PIPECONTROL command with "LRI
2091 * Post Sync Operation" in GPGPU mode of operation (i.e when
2092 * PIPELINE_SELECT command is set to GPGPU mode of operation)."
2094 * The same text exists a few rows below for Post Sync Op.
2096 * On Gen12 this is GEN:BUG:1607156449.
2098 if (bits
& ANV_PIPE_POST_SYNC_BIT
) {
2099 if ((GEN_GEN
== 9 || (GEN_GEN
== 12 && devinfo
->revision
== 0 /* A0 */)) &&
2100 cmd_buffer
->state
.current_pipeline
== GPGPU
)
2101 bits
|= ANV_PIPE_CS_STALL_BIT
;
2102 bits
&= ~ANV_PIPE_POST_SYNC_BIT
;
2105 if (bits
& (ANV_PIPE_FLUSH_BITS
| ANV_PIPE_CS_STALL_BIT
|
2106 ANV_PIPE_END_OF_PIPE_SYNC_BIT
)) {
2107 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pipe
) {
2109 pipe
.TileCacheFlushEnable
= bits
& ANV_PIPE_TILE_CACHE_FLUSH_BIT
;
2111 pipe
.DepthCacheFlushEnable
= bits
& ANV_PIPE_DEPTH_CACHE_FLUSH_BIT
;
2112 pipe
.DCFlushEnable
= bits
& ANV_PIPE_DATA_CACHE_FLUSH_BIT
;
2113 pipe
.RenderTargetCacheFlushEnable
=
2114 bits
& ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT
;
2116 /* GEN:BUG:1409600907: "PIPE_CONTROL with Depth Stall Enable bit must
2117 * be set with any PIPE_CONTROL with Depth Flush Enable bit set.
2120 pipe
.DepthStallEnable
=
2121 pipe
.DepthCacheFlushEnable
|| (bits
& ANV_PIPE_DEPTH_STALL_BIT
);
2123 pipe
.DepthStallEnable
= bits
& ANV_PIPE_DEPTH_STALL_BIT
;
2126 pipe
.CommandStreamerStallEnable
= bits
& ANV_PIPE_CS_STALL_BIT
;
2127 pipe
.StallAtPixelScoreboard
= bits
& ANV_PIPE_STALL_AT_SCOREBOARD_BIT
;
2129 /* From Sandybridge PRM, volume 2, "1.7.3.1 Writing a Value to Memory":
2131 * "The most common action to perform upon reaching a
2132 * synchronization point is to write a value out to memory. An
2133 * immediate value (included with the synchronization command) may
2137 * From Broadwell PRM, volume 7, "End-of-Pipe Synchronization":
2139 * "In case the data flushed out by the render engine is to be
2140 * read back in to the render engine in coherent manner, then the
2141 * render engine has to wait for the fence completion before
2142 * accessing the flushed data. This can be achieved by following
2143 * means on various products: PIPE_CONTROL command with CS Stall
2144 * and the required write caches flushed with Post-Sync-Operation
2145 * as Write Immediate Data.
2148 * - Workload-1 (3D/GPGPU/MEDIA)
2149 * - PIPE_CONTROL (CS Stall, Post-Sync-Operation Write
2150 * Immediate Data, Required Write Cache Flush bits set)
2151 * - Workload-2 (Can use the data produce or output by
2154 if (bits
& ANV_PIPE_END_OF_PIPE_SYNC_BIT
) {
2155 pipe
.CommandStreamerStallEnable
= true;
2156 pipe
.PostSyncOperation
= WriteImmediateData
;
2157 pipe
.Address
= (struct anv_address
) {
2158 .bo
= cmd_buffer
->device
->workaround_bo
,
2164 * According to the Broadwell documentation, any PIPE_CONTROL with the
2165 * "Command Streamer Stall" bit set must also have another bit set,
2166 * with five different options:
2168 * - Render Target Cache Flush
2169 * - Depth Cache Flush
2170 * - Stall at Pixel Scoreboard
2171 * - Post-Sync Operation
2175 * I chose "Stall at Pixel Scoreboard" since that's what we use in
2176 * mesa and it seems to work fine. The choice is fairly arbitrary.
2178 if (pipe
.CommandStreamerStallEnable
&&
2179 !pipe
.RenderTargetCacheFlushEnable
&&
2180 !pipe
.DepthCacheFlushEnable
&&
2181 !pipe
.StallAtPixelScoreboard
&&
2182 !pipe
.PostSyncOperation
&&
2183 !pipe
.DepthStallEnable
&&
2184 !pipe
.DCFlushEnable
)
2185 pipe
.StallAtPixelScoreboard
= true;
2188 /* If a render target flush was emitted, then we can toggle off the bit
2189 * saying that render target writes are ongoing.
2191 if (bits
& ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT
)
2192 bits
&= ~(ANV_PIPE_RENDER_TARGET_BUFFER_WRITES
);
2194 if (GEN_IS_HASWELL
) {
2195 /* Haswell needs addition work-arounds:
2197 * From Haswell PRM, volume 2, part 1, "End-of-Pipe Synchronization":
2200 * PIPE_CONTROL command with the CS Stall and the required write
2201 * caches flushed with Post-SyncOperation as Write Immediate Data
2202 * followed by eight dummy MI_STORE_DATA_IMM (write to scratch
2207 * - PIPE_CONTROL (CS Stall, Post-Sync-Operation Write
2208 * Immediate Data, Required Write Cache Flush bits set)
2209 * - MI_STORE_DATA_IMM (8 times) (Dummy data, Scratch Address)
2210 * - Workload-2 (Can use the data produce or output by
2213 * Unfortunately, both the PRMs and the internal docs are a bit
2214 * out-of-date in this regard. What the windows driver does (and
2215 * this appears to actually work) is to emit a register read from the
2216 * memory address written by the pipe control above.
2218 * What register we load into doesn't matter. We choose an indirect
2219 * rendering register because we know it always exists and it's one
2220 * of the first registers the command parser allows us to write. If
2221 * you don't have command parser support in your kernel (pre-4.2),
2222 * this will get turned into MI_NOOP and you won't get the
2223 * workaround. Unfortunately, there's just not much we can do in
2224 * that case. This register is perfectly safe to write since we
2225 * always re-load all of the indirect draw registers right before
2226 * 3DPRIMITIVE when needed anyway.
2228 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
2229 lrm
.RegisterAddress
= 0x243C; /* GEN7_3DPRIM_START_INSTANCE */
2230 lrm
.MemoryAddress
= (struct anv_address
) {
2231 .bo
= cmd_buffer
->device
->workaround_bo
,
2237 bits
&= ~(ANV_PIPE_FLUSH_BITS
| ANV_PIPE_CS_STALL_BIT
|
2238 ANV_PIPE_END_OF_PIPE_SYNC_BIT
);
2241 if (bits
& ANV_PIPE_INVALIDATE_BITS
) {
2242 /* From the SKL PRM, Vol. 2a, "PIPE_CONTROL",
2244 * "If the VF Cache Invalidation Enable is set to a 1 in a
2245 * PIPE_CONTROL, a separate Null PIPE_CONTROL, all bitfields sets to
2246 * 0, with the VF Cache Invalidation Enable set to 0 needs to be sent
2247 * prior to the PIPE_CONTROL with VF Cache Invalidation Enable set to
2250 * This appears to hang Broadwell, so we restrict it to just gen9.
2252 if (GEN_GEN
== 9 && (bits
& ANV_PIPE_VF_CACHE_INVALIDATE_BIT
))
2253 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pipe
);
2255 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pipe
) {
2256 pipe
.StateCacheInvalidationEnable
=
2257 bits
& ANV_PIPE_STATE_CACHE_INVALIDATE_BIT
;
2258 pipe
.ConstantCacheInvalidationEnable
=
2259 bits
& ANV_PIPE_CONSTANT_CACHE_INVALIDATE_BIT
;
2260 pipe
.VFCacheInvalidationEnable
=
2261 bits
& ANV_PIPE_VF_CACHE_INVALIDATE_BIT
;
2262 pipe
.TextureCacheInvalidationEnable
=
2263 bits
& ANV_PIPE_TEXTURE_CACHE_INVALIDATE_BIT
;
2264 pipe
.InstructionCacheInvalidateEnable
=
2265 bits
& ANV_PIPE_INSTRUCTION_CACHE_INVALIDATE_BIT
;
2267 /* From the SKL PRM, Vol. 2a, "PIPE_CONTROL",
2269 * "When VF Cache Invalidate is set “Post Sync Operation” must be
2270 * enabled to “Write Immediate Data” or “Write PS Depth Count” or
2271 * “Write Timestamp”.
2273 if (GEN_GEN
== 9 && pipe
.VFCacheInvalidationEnable
) {
2274 pipe
.PostSyncOperation
= WriteImmediateData
;
2276 (struct anv_address
) { cmd_buffer
->device
->workaround_bo
, 0 };
2281 if ((bits
& ANV_PIPE_AUX_TABLE_INVALIDATE_BIT
) &&
2282 cmd_buffer
->device
->info
.has_aux_map
) {
2283 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_LOAD_REGISTER_IMM
), lri
) {
2284 lri
.RegisterOffset
= GENX(GFX_CCS_AUX_INV_num
);
2290 bits
&= ~ANV_PIPE_INVALIDATE_BITS
;
2293 cmd_buffer
->state
.pending_pipe_bits
= bits
;
2296 void genX(CmdPipelineBarrier
)(
2297 VkCommandBuffer commandBuffer
,
2298 VkPipelineStageFlags srcStageMask
,
2299 VkPipelineStageFlags destStageMask
,
2301 uint32_t memoryBarrierCount
,
2302 const VkMemoryBarrier
* pMemoryBarriers
,
2303 uint32_t bufferMemoryBarrierCount
,
2304 const VkBufferMemoryBarrier
* pBufferMemoryBarriers
,
2305 uint32_t imageMemoryBarrierCount
,
2306 const VkImageMemoryBarrier
* pImageMemoryBarriers
)
2308 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2310 /* XXX: Right now, we're really dumb and just flush whatever categories
2311 * the app asks for. One of these days we may make this a bit better
2312 * but right now that's all the hardware allows for in most areas.
2314 VkAccessFlags src_flags
= 0;
2315 VkAccessFlags dst_flags
= 0;
2317 for (uint32_t i
= 0; i
< memoryBarrierCount
; i
++) {
2318 src_flags
|= pMemoryBarriers
[i
].srcAccessMask
;
2319 dst_flags
|= pMemoryBarriers
[i
].dstAccessMask
;
2322 for (uint32_t i
= 0; i
< bufferMemoryBarrierCount
; i
++) {
2323 src_flags
|= pBufferMemoryBarriers
[i
].srcAccessMask
;
2324 dst_flags
|= pBufferMemoryBarriers
[i
].dstAccessMask
;
2327 for (uint32_t i
= 0; i
< imageMemoryBarrierCount
; i
++) {
2328 src_flags
|= pImageMemoryBarriers
[i
].srcAccessMask
;
2329 dst_flags
|= pImageMemoryBarriers
[i
].dstAccessMask
;
2330 ANV_FROM_HANDLE(anv_image
, image
, pImageMemoryBarriers
[i
].image
);
2331 const VkImageSubresourceRange
*range
=
2332 &pImageMemoryBarriers
[i
].subresourceRange
;
2334 uint32_t base_layer
, layer_count
;
2335 if (image
->type
== VK_IMAGE_TYPE_3D
) {
2337 layer_count
= anv_minify(image
->extent
.depth
, range
->baseMipLevel
);
2339 base_layer
= range
->baseArrayLayer
;
2340 layer_count
= anv_get_layerCount(image
, range
);
2343 if (range
->aspectMask
& VK_IMAGE_ASPECT_DEPTH_BIT
) {
2344 transition_depth_buffer(cmd_buffer
, image
,
2345 pImageMemoryBarriers
[i
].oldLayout
,
2346 pImageMemoryBarriers
[i
].newLayout
);
2349 if (range
->aspectMask
& VK_IMAGE_ASPECT_STENCIL_BIT
) {
2350 transition_stencil_buffer(cmd_buffer
, image
,
2351 range
->baseMipLevel
,
2352 anv_get_levelCount(image
, range
),
2353 base_layer
, layer_count
,
2354 pImageMemoryBarriers
[i
].oldLayout
,
2355 pImageMemoryBarriers
[i
].newLayout
);
2358 if (range
->aspectMask
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) {
2359 VkImageAspectFlags color_aspects
=
2360 anv_image_expand_aspects(image
, range
->aspectMask
);
2361 uint32_t aspect_bit
;
2362 anv_foreach_image_aspect_bit(aspect_bit
, image
, color_aspects
) {
2363 transition_color_buffer(cmd_buffer
, image
, 1UL << aspect_bit
,
2364 range
->baseMipLevel
,
2365 anv_get_levelCount(image
, range
),
2366 base_layer
, layer_count
,
2367 pImageMemoryBarriers
[i
].oldLayout
,
2368 pImageMemoryBarriers
[i
].newLayout
);
2373 cmd_buffer
->state
.pending_pipe_bits
|=
2374 anv_pipe_flush_bits_for_access_flags(src_flags
) |
2375 anv_pipe_invalidate_bits_for_access_flags(dst_flags
);
2379 cmd_buffer_alloc_push_constants(struct anv_cmd_buffer
*cmd_buffer
)
2381 VkShaderStageFlags stages
=
2382 cmd_buffer
->state
.gfx
.base
.pipeline
->active_stages
;
2384 /* In order to avoid thrash, we assume that vertex and fragment stages
2385 * always exist. In the rare case where one is missing *and* the other
2386 * uses push concstants, this may be suboptimal. However, avoiding stalls
2387 * seems more important.
2389 stages
|= VK_SHADER_STAGE_FRAGMENT_BIT
| VK_SHADER_STAGE_VERTEX_BIT
;
2391 if (stages
== cmd_buffer
->state
.push_constant_stages
)
2395 const unsigned push_constant_kb
= 32;
2396 #elif GEN_IS_HASWELL
2397 const unsigned push_constant_kb
= cmd_buffer
->device
->info
.gt
== 3 ? 32 : 16;
2399 const unsigned push_constant_kb
= 16;
2402 const unsigned num_stages
=
2403 util_bitcount(stages
& VK_SHADER_STAGE_ALL_GRAPHICS
);
2404 unsigned size_per_stage
= push_constant_kb
/ num_stages
;
2406 /* Broadwell+ and Haswell gt3 require that the push constant sizes be in
2407 * units of 2KB. Incidentally, these are the same platforms that have
2408 * 32KB worth of push constant space.
2410 if (push_constant_kb
== 32)
2411 size_per_stage
&= ~1u;
2413 uint32_t kb_used
= 0;
2414 for (int i
= MESA_SHADER_VERTEX
; i
< MESA_SHADER_FRAGMENT
; i
++) {
2415 unsigned push_size
= (stages
& (1 << i
)) ? size_per_stage
: 0;
2416 anv_batch_emit(&cmd_buffer
->batch
,
2417 GENX(3DSTATE_PUSH_CONSTANT_ALLOC_VS
), alloc
) {
2418 alloc
._3DCommandSubOpcode
= 18 + i
;
2419 alloc
.ConstantBufferOffset
= (push_size
> 0) ? kb_used
: 0;
2420 alloc
.ConstantBufferSize
= push_size
;
2422 kb_used
+= push_size
;
2425 anv_batch_emit(&cmd_buffer
->batch
,
2426 GENX(3DSTATE_PUSH_CONSTANT_ALLOC_PS
), alloc
) {
2427 alloc
.ConstantBufferOffset
= kb_used
;
2428 alloc
.ConstantBufferSize
= push_constant_kb
- kb_used
;
2431 cmd_buffer
->state
.push_constant_stages
= stages
;
2433 /* From the BDW PRM for 3DSTATE_PUSH_CONSTANT_ALLOC_VS:
2435 * "The 3DSTATE_CONSTANT_VS must be reprogrammed prior to
2436 * the next 3DPRIMITIVE command after programming the
2437 * 3DSTATE_PUSH_CONSTANT_ALLOC_VS"
2439 * Since 3DSTATE_PUSH_CONSTANT_ALLOC_VS is programmed as part of
2440 * pipeline setup, we need to dirty push constants.
2442 cmd_buffer
->state
.push_constants_dirty
|= VK_SHADER_STAGE_ALL_GRAPHICS
;
2445 static struct anv_address
2446 anv_descriptor_set_address(struct anv_cmd_buffer
*cmd_buffer
,
2447 struct anv_descriptor_set
*set
)
2450 /* This is a normal descriptor set */
2451 return (struct anv_address
) {
2452 .bo
= set
->pool
->bo
,
2453 .offset
= set
->desc_mem
.offset
,
2456 /* This is a push descriptor set. We have to flag it as used on the GPU
2457 * so that the next time we push descriptors, we grab a new memory.
2459 struct anv_push_descriptor_set
*push_set
=
2460 (struct anv_push_descriptor_set
*)set
;
2461 push_set
->set_used_on_gpu
= true;
2463 return (struct anv_address
) {
2464 .bo
= cmd_buffer
->dynamic_state_stream
.state_pool
->block_pool
.bo
,
2465 .offset
= set
->desc_mem
.offset
,
2470 static struct anv_cmd_pipeline_state
*
2471 pipe_state_for_stage(struct anv_cmd_buffer
*cmd_buffer
,
2472 gl_shader_stage stage
)
2475 case MESA_SHADER_COMPUTE
:
2476 return &cmd_buffer
->state
.compute
.base
;
2478 case MESA_SHADER_VERTEX
:
2479 case MESA_SHADER_TESS_CTRL
:
2480 case MESA_SHADER_TESS_EVAL
:
2481 case MESA_SHADER_GEOMETRY
:
2482 case MESA_SHADER_FRAGMENT
:
2483 return &cmd_buffer
->state
.gfx
.base
;
2486 unreachable("invalid stage");
2491 emit_binding_table(struct anv_cmd_buffer
*cmd_buffer
,
2492 gl_shader_stage stage
,
2493 struct anv_state
*bt_state
)
2495 struct anv_subpass
*subpass
= cmd_buffer
->state
.subpass
;
2496 uint32_t state_offset
;
2498 struct anv_cmd_pipeline_state
*pipe_state
=
2499 pipe_state_for_stage(cmd_buffer
, stage
);
2500 struct anv_pipeline
*pipeline
= pipe_state
->pipeline
;
2502 if (!anv_pipeline_has_stage(pipeline
, stage
)) {
2503 *bt_state
= (struct anv_state
) { 0, };
2507 struct anv_pipeline_bind_map
*map
= &pipeline
->shaders
[stage
]->bind_map
;
2508 if (map
->surface_count
== 0) {
2509 *bt_state
= (struct anv_state
) { 0, };
2513 *bt_state
= anv_cmd_buffer_alloc_binding_table(cmd_buffer
,
2516 uint32_t *bt_map
= bt_state
->map
;
2518 if (bt_state
->map
== NULL
)
2519 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2521 /* We only need to emit relocs if we're not using softpin. If we are using
2522 * softpin then we always keep all user-allocated memory objects resident.
2524 const bool need_client_mem_relocs
=
2525 !cmd_buffer
->device
->physical
->use_softpin
;
2527 for (uint32_t s
= 0; s
< map
->surface_count
; s
++) {
2528 struct anv_pipeline_binding
*binding
= &map
->surface_to_descriptor
[s
];
2530 struct anv_state surface_state
;
2532 switch (binding
->set
) {
2533 case ANV_DESCRIPTOR_SET_NULL
:
2537 case ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS
:
2538 /* Color attachment binding */
2539 assert(stage
== MESA_SHADER_FRAGMENT
);
2540 if (binding
->index
< subpass
->color_count
) {
2541 const unsigned att
=
2542 subpass
->color_attachments
[binding
->index
].attachment
;
2544 /* From the Vulkan 1.0.46 spec:
2546 * "If any color or depth/stencil attachments are
2547 * VK_ATTACHMENT_UNUSED, then no writes occur for those
2550 if (att
== VK_ATTACHMENT_UNUSED
) {
2551 surface_state
= cmd_buffer
->state
.null_surface_state
;
2553 surface_state
= cmd_buffer
->state
.attachments
[att
].color
.state
;
2556 surface_state
= cmd_buffer
->state
.null_surface_state
;
2559 bt_map
[s
] = surface_state
.offset
+ state_offset
;
2562 case ANV_DESCRIPTOR_SET_SHADER_CONSTANTS
: {
2563 struct anv_state surface_state
=
2564 anv_cmd_buffer_alloc_surface_state(cmd_buffer
);
2566 struct anv_address constant_data
= {
2567 .bo
= pipeline
->device
->dynamic_state_pool
.block_pool
.bo
,
2568 .offset
= pipeline
->shaders
[stage
]->constant_data
.offset
,
2570 unsigned constant_data_size
=
2571 pipeline
->shaders
[stage
]->constant_data_size
;
2573 const enum isl_format format
=
2574 anv_isl_format_for_descriptor_type(VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER
);
2575 anv_fill_buffer_surface_state(cmd_buffer
->device
,
2576 surface_state
, format
,
2577 constant_data
, constant_data_size
, 1);
2579 bt_map
[s
] = surface_state
.offset
+ state_offset
;
2580 add_surface_reloc(cmd_buffer
, surface_state
, constant_data
);
2584 case ANV_DESCRIPTOR_SET_NUM_WORK_GROUPS
: {
2585 /* This is always the first binding for compute shaders */
2586 assert(stage
== MESA_SHADER_COMPUTE
&& s
== 0);
2588 struct anv_state surface_state
=
2589 anv_cmd_buffer_alloc_surface_state(cmd_buffer
);
2591 const enum isl_format format
=
2592 anv_isl_format_for_descriptor_type(VK_DESCRIPTOR_TYPE_STORAGE_BUFFER
);
2593 anv_fill_buffer_surface_state(cmd_buffer
->device
, surface_state
,
2595 cmd_buffer
->state
.compute
.num_workgroups
,
2597 bt_map
[s
] = surface_state
.offset
+ state_offset
;
2598 if (need_client_mem_relocs
) {
2599 add_surface_reloc(cmd_buffer
, surface_state
,
2600 cmd_buffer
->state
.compute
.num_workgroups
);
2605 case ANV_DESCRIPTOR_SET_DESCRIPTORS
: {
2606 /* This is a descriptor set buffer so the set index is actually
2607 * given by binding->binding. (Yes, that's confusing.)
2609 struct anv_descriptor_set
*set
=
2610 pipe_state
->descriptors
[binding
->index
];
2611 assert(set
->desc_mem
.alloc_size
);
2612 assert(set
->desc_surface_state
.alloc_size
);
2613 bt_map
[s
] = set
->desc_surface_state
.offset
+ state_offset
;
2614 add_surface_reloc(cmd_buffer
, set
->desc_surface_state
,
2615 anv_descriptor_set_address(cmd_buffer
, set
));
2620 assert(binding
->set
< MAX_SETS
);
2621 const struct anv_descriptor
*desc
=
2622 &pipe_state
->descriptors
[binding
->set
]->descriptors
[binding
->index
];
2624 switch (desc
->type
) {
2625 case VK_DESCRIPTOR_TYPE_SAMPLER
:
2626 /* Nothing for us to do here */
2629 case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER
:
2630 case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE
: {
2631 struct anv_surface_state sstate
=
2632 (desc
->layout
== VK_IMAGE_LAYOUT_GENERAL
) ?
2633 desc
->image_view
->planes
[binding
->plane
].general_sampler_surface_state
:
2634 desc
->image_view
->planes
[binding
->plane
].optimal_sampler_surface_state
;
2635 surface_state
= sstate
.state
;
2636 assert(surface_state
.alloc_size
);
2637 if (need_client_mem_relocs
)
2638 add_surface_state_relocs(cmd_buffer
, sstate
);
2641 case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT
:
2642 assert(stage
== MESA_SHADER_FRAGMENT
);
2643 if ((desc
->image_view
->aspect_mask
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) == 0) {
2644 /* For depth and stencil input attachments, we treat it like any
2645 * old texture that a user may have bound.
2647 assert(desc
->image_view
->n_planes
== 1);
2648 struct anv_surface_state sstate
=
2649 (desc
->layout
== VK_IMAGE_LAYOUT_GENERAL
) ?
2650 desc
->image_view
->planes
[0].general_sampler_surface_state
:
2651 desc
->image_view
->planes
[0].optimal_sampler_surface_state
;
2652 surface_state
= sstate
.state
;
2653 assert(surface_state
.alloc_size
);
2654 if (need_client_mem_relocs
)
2655 add_surface_state_relocs(cmd_buffer
, sstate
);
2657 /* For color input attachments, we create the surface state at
2658 * vkBeginRenderPass time so that we can include aux and clear
2659 * color information.
2661 assert(binding
->input_attachment_index
< subpass
->input_count
);
2662 const unsigned subpass_att
= binding
->input_attachment_index
;
2663 const unsigned att
= subpass
->input_attachments
[subpass_att
].attachment
;
2664 surface_state
= cmd_buffer
->state
.attachments
[att
].input
.state
;
2668 case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE
: {
2669 struct anv_surface_state sstate
= (binding
->write_only
)
2670 ? desc
->image_view
->planes
[binding
->plane
].writeonly_storage_surface_state
2671 : desc
->image_view
->planes
[binding
->plane
].storage_surface_state
;
2672 surface_state
= sstate
.state
;
2673 assert(surface_state
.alloc_size
);
2674 if (need_client_mem_relocs
)
2675 add_surface_state_relocs(cmd_buffer
, sstate
);
2679 case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER
:
2680 case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER
:
2681 case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER
:
2682 surface_state
= desc
->buffer_view
->surface_state
;
2683 assert(surface_state
.alloc_size
);
2684 if (need_client_mem_relocs
) {
2685 add_surface_reloc(cmd_buffer
, surface_state
,
2686 desc
->buffer_view
->address
);
2690 case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC
:
2691 case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC
: {
2692 /* Compute the offset within the buffer */
2693 struct anv_push_constants
*push
=
2694 &cmd_buffer
->state
.push_constants
[stage
];
2696 uint32_t dynamic_offset
=
2697 push
->dynamic_offsets
[binding
->dynamic_offset_index
];
2698 uint64_t offset
= desc
->offset
+ dynamic_offset
;
2699 /* Clamp to the buffer size */
2700 offset
= MIN2(offset
, desc
->buffer
->size
);
2701 /* Clamp the range to the buffer size */
2702 uint32_t range
= MIN2(desc
->range
, desc
->buffer
->size
- offset
);
2704 struct anv_address address
=
2705 anv_address_add(desc
->buffer
->address
, offset
);
2708 anv_state_stream_alloc(&cmd_buffer
->surface_state_stream
, 64, 64);
2709 enum isl_format format
=
2710 anv_isl_format_for_descriptor_type(desc
->type
);
2712 anv_fill_buffer_surface_state(cmd_buffer
->device
, surface_state
,
2713 format
, address
, range
, 1);
2714 if (need_client_mem_relocs
)
2715 add_surface_reloc(cmd_buffer
, surface_state
, address
);
2719 case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER
:
2720 surface_state
= (binding
->write_only
)
2721 ? desc
->buffer_view
->writeonly_storage_surface_state
2722 : desc
->buffer_view
->storage_surface_state
;
2723 assert(surface_state
.alloc_size
);
2724 if (need_client_mem_relocs
) {
2725 add_surface_reloc(cmd_buffer
, surface_state
,
2726 desc
->buffer_view
->address
);
2731 assert(!"Invalid descriptor type");
2734 bt_map
[s
] = surface_state
.offset
+ state_offset
;
2744 emit_samplers(struct anv_cmd_buffer
*cmd_buffer
,
2745 gl_shader_stage stage
,
2746 struct anv_state
*state
)
2748 struct anv_cmd_pipeline_state
*pipe_state
=
2749 pipe_state_for_stage(cmd_buffer
, stage
);
2750 struct anv_pipeline
*pipeline
= pipe_state
->pipeline
;
2752 if (!anv_pipeline_has_stage(pipeline
, stage
)) {
2753 *state
= (struct anv_state
) { 0, };
2757 struct anv_pipeline_bind_map
*map
= &pipeline
->shaders
[stage
]->bind_map
;
2758 if (map
->sampler_count
== 0) {
2759 *state
= (struct anv_state
) { 0, };
2763 uint32_t size
= map
->sampler_count
* 16;
2764 *state
= anv_cmd_buffer_alloc_dynamic_state(cmd_buffer
, size
, 32);
2766 if (state
->map
== NULL
)
2767 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2769 for (uint32_t s
= 0; s
< map
->sampler_count
; s
++) {
2770 struct anv_pipeline_binding
*binding
= &map
->sampler_to_descriptor
[s
];
2771 const struct anv_descriptor
*desc
=
2772 &pipe_state
->descriptors
[binding
->set
]->descriptors
[binding
->index
];
2774 if (desc
->type
!= VK_DESCRIPTOR_TYPE_SAMPLER
&&
2775 desc
->type
!= VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER
)
2778 struct anv_sampler
*sampler
= desc
->sampler
;
2780 /* This can happen if we have an unfilled slot since TYPE_SAMPLER
2781 * happens to be zero.
2783 if (sampler
== NULL
)
2786 memcpy(state
->map
+ (s
* 16),
2787 sampler
->state
[binding
->plane
], sizeof(sampler
->state
[0]));
2794 flush_descriptor_sets(struct anv_cmd_buffer
*cmd_buffer
,
2795 struct anv_pipeline
*pipeline
)
2797 VkShaderStageFlags dirty
= cmd_buffer
->state
.descriptors_dirty
&
2798 pipeline
->active_stages
;
2800 VkResult result
= VK_SUCCESS
;
2801 anv_foreach_stage(s
, dirty
) {
2802 result
= emit_samplers(cmd_buffer
, s
, &cmd_buffer
->state
.samplers
[s
]);
2803 if (result
!= VK_SUCCESS
)
2805 result
= emit_binding_table(cmd_buffer
, s
,
2806 &cmd_buffer
->state
.binding_tables
[s
]);
2807 if (result
!= VK_SUCCESS
)
2811 if (result
!= VK_SUCCESS
) {
2812 assert(result
== VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2814 result
= anv_cmd_buffer_new_binding_table_block(cmd_buffer
);
2815 if (result
!= VK_SUCCESS
)
2818 /* Re-emit state base addresses so we get the new surface state base
2819 * address before we start emitting binding tables etc.
2821 genX(cmd_buffer_emit_state_base_address
)(cmd_buffer
);
2823 /* Re-emit all active binding tables */
2824 dirty
|= pipeline
->active_stages
;
2825 anv_foreach_stage(s
, dirty
) {
2826 result
= emit_samplers(cmd_buffer
, s
, &cmd_buffer
->state
.samplers
[s
]);
2827 if (result
!= VK_SUCCESS
) {
2828 anv_batch_set_error(&cmd_buffer
->batch
, result
);
2831 result
= emit_binding_table(cmd_buffer
, s
,
2832 &cmd_buffer
->state
.binding_tables
[s
]);
2833 if (result
!= VK_SUCCESS
) {
2834 anv_batch_set_error(&cmd_buffer
->batch
, result
);
2840 cmd_buffer
->state
.descriptors_dirty
&= ~dirty
;
2846 cmd_buffer_emit_descriptor_pointers(struct anv_cmd_buffer
*cmd_buffer
,
2849 static const uint32_t sampler_state_opcodes
[] = {
2850 [MESA_SHADER_VERTEX
] = 43,
2851 [MESA_SHADER_TESS_CTRL
] = 44, /* HS */
2852 [MESA_SHADER_TESS_EVAL
] = 45, /* DS */
2853 [MESA_SHADER_GEOMETRY
] = 46,
2854 [MESA_SHADER_FRAGMENT
] = 47,
2855 [MESA_SHADER_COMPUTE
] = 0,
2858 static const uint32_t binding_table_opcodes
[] = {
2859 [MESA_SHADER_VERTEX
] = 38,
2860 [MESA_SHADER_TESS_CTRL
] = 39,
2861 [MESA_SHADER_TESS_EVAL
] = 40,
2862 [MESA_SHADER_GEOMETRY
] = 41,
2863 [MESA_SHADER_FRAGMENT
] = 42,
2864 [MESA_SHADER_COMPUTE
] = 0,
2867 anv_foreach_stage(s
, stages
) {
2868 assert(s
< ARRAY_SIZE(binding_table_opcodes
));
2869 assert(binding_table_opcodes
[s
] > 0);
2871 if (cmd_buffer
->state
.samplers
[s
].alloc_size
> 0) {
2872 anv_batch_emit(&cmd_buffer
->batch
,
2873 GENX(3DSTATE_SAMPLER_STATE_POINTERS_VS
), ssp
) {
2874 ssp
._3DCommandSubOpcode
= sampler_state_opcodes
[s
];
2875 ssp
.PointertoVSSamplerState
= cmd_buffer
->state
.samplers
[s
].offset
;
2879 /* Always emit binding table pointers if we're asked to, since on SKL
2880 * this is what flushes push constants. */
2881 anv_batch_emit(&cmd_buffer
->batch
,
2882 GENX(3DSTATE_BINDING_TABLE_POINTERS_VS
), btp
) {
2883 btp
._3DCommandSubOpcode
= binding_table_opcodes
[s
];
2884 btp
.PointertoVSBindingTable
= cmd_buffer
->state
.binding_tables
[s
].offset
;
2889 static struct anv_address
2890 get_push_range_address(struct anv_cmd_buffer
*cmd_buffer
,
2891 gl_shader_stage stage
,
2892 const struct anv_push_range
*range
)
2894 const struct anv_cmd_graphics_state
*gfx_state
= &cmd_buffer
->state
.gfx
;
2895 switch (range
->set
) {
2896 case ANV_DESCRIPTOR_SET_DESCRIPTORS
: {
2897 /* This is a descriptor set buffer so the set index is
2898 * actually given by binding->binding. (Yes, that's
2901 struct anv_descriptor_set
*set
=
2902 gfx_state
->base
.descriptors
[range
->index
];
2903 return anv_descriptor_set_address(cmd_buffer
, set
);
2907 case ANV_DESCRIPTOR_SET_PUSH_CONSTANTS
: {
2908 struct anv_state state
=
2909 anv_cmd_buffer_push_constants(cmd_buffer
, stage
);
2910 return (struct anv_address
) {
2911 .bo
= cmd_buffer
->device
->dynamic_state_pool
.block_pool
.bo
,
2912 .offset
= state
.offset
,
2918 assert(range
->set
< MAX_SETS
);
2919 struct anv_descriptor_set
*set
=
2920 gfx_state
->base
.descriptors
[range
->set
];
2921 const struct anv_descriptor
*desc
=
2922 &set
->descriptors
[range
->index
];
2924 if (desc
->type
== VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER
) {
2925 return desc
->buffer_view
->address
;
2927 assert(desc
->type
== VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC
);
2928 struct anv_push_constants
*push
=
2929 &cmd_buffer
->state
.push_constants
[stage
];
2930 uint32_t dynamic_offset
=
2931 push
->dynamic_offsets
[range
->dynamic_offset_index
];
2932 return anv_address_add(desc
->buffer
->address
,
2933 desc
->offset
+ dynamic_offset
);
2940 cmd_buffer_emit_push_constant(struct anv_cmd_buffer
*cmd_buffer
,
2941 gl_shader_stage stage
,
2942 struct anv_address
*buffers
,
2943 unsigned buffer_count
)
2945 const struct anv_cmd_graphics_state
*gfx_state
= &cmd_buffer
->state
.gfx
;
2946 const struct anv_pipeline
*pipeline
= gfx_state
->base
.pipeline
;
2948 static const uint32_t push_constant_opcodes
[] = {
2949 [MESA_SHADER_VERTEX
] = 21,
2950 [MESA_SHADER_TESS_CTRL
] = 25, /* HS */
2951 [MESA_SHADER_TESS_EVAL
] = 26, /* DS */
2952 [MESA_SHADER_GEOMETRY
] = 22,
2953 [MESA_SHADER_FRAGMENT
] = 23,
2954 [MESA_SHADER_COMPUTE
] = 0,
2957 assert(stage
< ARRAY_SIZE(push_constant_opcodes
));
2958 assert(push_constant_opcodes
[stage
] > 0);
2960 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DSTATE_CONSTANT_VS
), c
) {
2961 c
._3DCommandSubOpcode
= push_constant_opcodes
[stage
];
2963 if (anv_pipeline_has_stage(pipeline
, stage
)) {
2964 const struct anv_pipeline_bind_map
*bind_map
=
2965 &pipeline
->shaders
[stage
]->bind_map
;
2968 c
.MOCS
= cmd_buffer
->device
->isl_dev
.mocs
.internal
;
2971 #if GEN_GEN >= 8 || GEN_IS_HASWELL
2972 /* The Skylake PRM contains the following restriction:
2974 * "The driver must ensure The following case does not occur
2975 * without a flush to the 3D engine: 3DSTATE_CONSTANT_* with
2976 * buffer 3 read length equal to zero committed followed by a
2977 * 3DSTATE_CONSTANT_* with buffer 0 read length not equal to
2980 * To avoid this, we program the buffers in the highest slots.
2981 * This way, slot 0 is only used if slot 3 is also used.
2983 assert(buffer_count
<= 4);
2984 const unsigned shift
= 4 - buffer_count
;
2985 for (unsigned i
= 0; i
< buffer_count
; i
++) {
2986 const struct anv_push_range
*range
= &bind_map
->push_ranges
[i
];
2988 /* At this point we only have non-empty ranges */
2989 assert(range
->length
> 0);
2991 /* For Ivy Bridge, make sure we only set the first range (actual
2994 assert((GEN_GEN
>= 8 || GEN_IS_HASWELL
) || i
== 0);
2996 c
.ConstantBody
.ReadLength
[i
+ shift
] = range
->length
;
2997 c
.ConstantBody
.Buffer
[i
+ shift
] =
2998 anv_address_add(buffers
[i
], range
->start
* 32);
3001 /* For Ivy Bridge, push constants are relative to dynamic state
3002 * base address and we only ever push actual push constants.
3004 if (bind_map
->push_ranges
[0].length
> 0) {
3005 assert(buffer_count
== 1);
3006 assert(bind_map
->push_ranges
[0].set
==
3007 ANV_DESCRIPTOR_SET_PUSH_CONSTANTS
);
3008 assert(buffers
[0].bo
==
3009 cmd_buffer
->device
->dynamic_state_pool
.block_pool
.bo
);
3010 c
.ConstantBody
.ReadLength
[0] = bind_map
->push_ranges
[0].length
;
3011 c
.ConstantBody
.Buffer
[0].bo
= NULL
;
3012 c
.ConstantBody
.Buffer
[0].offset
= buffers
[0].offset
;
3014 assert(bind_map
->push_ranges
[1].length
== 0);
3015 assert(bind_map
->push_ranges
[2].length
== 0);
3016 assert(bind_map
->push_ranges
[3].length
== 0);
3024 cmd_buffer_emit_push_constant_all(struct anv_cmd_buffer
*cmd_buffer
,
3025 uint32_t shader_mask
,
3026 struct anv_address
*buffers
,
3027 uint32_t buffer_count
)
3029 if (buffer_count
== 0) {
3030 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DSTATE_CONSTANT_ALL
), c
) {
3031 c
.ShaderUpdateEnable
= shader_mask
;
3032 c
.MOCS
= cmd_buffer
->device
->isl_dev
.mocs
.internal
;
3037 const struct anv_cmd_graphics_state
*gfx_state
= &cmd_buffer
->state
.gfx
;
3038 const struct anv_pipeline
*pipeline
= gfx_state
->base
.pipeline
;
3040 static const uint32_t push_constant_opcodes
[] = {
3041 [MESA_SHADER_VERTEX
] = 21,
3042 [MESA_SHADER_TESS_CTRL
] = 25, /* HS */
3043 [MESA_SHADER_TESS_EVAL
] = 26, /* DS */
3044 [MESA_SHADER_GEOMETRY
] = 22,
3045 [MESA_SHADER_FRAGMENT
] = 23,
3046 [MESA_SHADER_COMPUTE
] = 0,
3049 gl_shader_stage stage
= vk_to_mesa_shader_stage(shader_mask
);
3050 assert(stage
< ARRAY_SIZE(push_constant_opcodes
));
3051 assert(push_constant_opcodes
[stage
] > 0);
3053 const struct anv_pipeline_bind_map
*bind_map
=
3054 &pipeline
->shaders
[stage
]->bind_map
;
3057 const uint32_t buffer_mask
= (1 << buffer_count
) - 1;
3058 const uint32_t num_dwords
= 2 + 2 * buffer_count
;
3060 dw
= anv_batch_emitn(&cmd_buffer
->batch
, num_dwords
,
3061 GENX(3DSTATE_CONSTANT_ALL
),
3062 .ShaderUpdateEnable
= shader_mask
,
3063 .PointerBufferMask
= buffer_mask
,
3064 .MOCS
= cmd_buffer
->device
->isl_dev
.mocs
.internal
);
3066 for (int i
= 0; i
< buffer_count
; i
++) {
3067 const struct anv_push_range
*range
= &bind_map
->push_ranges
[i
];
3068 GENX(3DSTATE_CONSTANT_ALL_DATA_pack
)(
3069 &cmd_buffer
->batch
, dw
+ 2 + i
* 2,
3070 &(struct GENX(3DSTATE_CONSTANT_ALL_DATA
)) {
3071 .PointerToConstantBuffer
=
3072 anv_address_add(buffers
[i
], range
->start
* 32),
3073 .ConstantBufferReadLength
= range
->length
,
3080 cmd_buffer_flush_push_constants(struct anv_cmd_buffer
*cmd_buffer
,
3081 VkShaderStageFlags dirty_stages
)
3083 VkShaderStageFlags flushed
= 0;
3084 const struct anv_cmd_graphics_state
*gfx_state
= &cmd_buffer
->state
.gfx
;
3085 const struct anv_pipeline
*pipeline
= gfx_state
->base
.pipeline
;
3088 uint32_t nobuffer_stages
= 0;
3091 anv_foreach_stage(stage
, dirty_stages
) {
3092 unsigned buffer_count
= 0;
3093 flushed
|= mesa_to_vk_shader_stage(stage
);
3094 UNUSED
uint32_t max_push_range
= 0;
3096 struct anv_address buffers
[4] = {};
3097 if (anv_pipeline_has_stage(pipeline
, stage
)) {
3098 const struct anv_pipeline_bind_map
*bind_map
=
3099 &pipeline
->shaders
[stage
]->bind_map
;
3101 for (unsigned i
= 0; i
< 4; i
++) {
3102 const struct anv_push_range
*range
= &bind_map
->push_ranges
[i
];
3103 if (range
->length
== 0)
3106 buffers
[i
] = get_push_range_address(cmd_buffer
, stage
, range
);
3107 max_push_range
= MAX2(max_push_range
, range
->length
);
3111 /* We have at most 4 buffers but they should be tightly packed */
3112 for (unsigned i
= buffer_count
; i
< 4; i
++)
3113 assert(bind_map
->push_ranges
[i
].length
== 0);
3117 /* If this stage doesn't have any push constants, emit it later in a
3118 * single CONSTANT_ALL packet.
3120 if (buffer_count
== 0) {
3121 nobuffer_stages
|= 1 << stage
;
3125 /* The Constant Buffer Read Length field from 3DSTATE_CONSTANT_ALL
3126 * contains only 5 bits, so we can only use it for buffers smaller than
3129 if (max_push_range
< 32) {
3130 cmd_buffer_emit_push_constant_all(cmd_buffer
, 1 << stage
,
3131 buffers
, buffer_count
);
3136 cmd_buffer_emit_push_constant(cmd_buffer
, stage
, buffers
, buffer_count
);
3140 if (nobuffer_stages
)
3141 cmd_buffer_emit_push_constant_all(cmd_buffer
, nobuffer_stages
, NULL
, 0);
3144 cmd_buffer
->state
.push_constants_dirty
&= ~flushed
;
3148 genX(cmd_buffer_flush_state
)(struct anv_cmd_buffer
*cmd_buffer
)
3150 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.gfx
.base
.pipeline
;
3153 uint32_t vb_emit
= cmd_buffer
->state
.gfx
.vb_dirty
& pipeline
->vb_used
;
3154 if (cmd_buffer
->state
.gfx
.dirty
& ANV_CMD_DIRTY_PIPELINE
)
3155 vb_emit
|= pipeline
->vb_used
;
3157 assert((pipeline
->active_stages
& VK_SHADER_STAGE_COMPUTE_BIT
) == 0);
3159 genX(cmd_buffer_config_l3
)(cmd_buffer
, pipeline
->l3_config
);
3161 genX(cmd_buffer_emit_hashing_mode
)(cmd_buffer
, UINT_MAX
, UINT_MAX
, 1);
3163 genX(flush_pipeline_select_3d
)(cmd_buffer
);
3166 const uint32_t num_buffers
= __builtin_popcount(vb_emit
);
3167 const uint32_t num_dwords
= 1 + num_buffers
* 4;
3169 p
= anv_batch_emitn(&cmd_buffer
->batch
, num_dwords
,
3170 GENX(3DSTATE_VERTEX_BUFFERS
));
3172 for_each_bit(vb
, vb_emit
) {
3173 struct anv_buffer
*buffer
= cmd_buffer
->state
.vertex_bindings
[vb
].buffer
;
3174 uint32_t offset
= cmd_buffer
->state
.vertex_bindings
[vb
].offset
;
3176 struct GENX(VERTEX_BUFFER_STATE
) state
= {
3177 .VertexBufferIndex
= vb
,
3179 .MOCS
= anv_mocs_for_bo(cmd_buffer
->device
, buffer
->address
.bo
),
3181 .BufferAccessType
= pipeline
->vb
[vb
].instanced
? INSTANCEDATA
: VERTEXDATA
,
3182 .InstanceDataStepRate
= pipeline
->vb
[vb
].instance_divisor
,
3185 .AddressModifyEnable
= true,
3186 .BufferPitch
= pipeline
->vb
[vb
].stride
,
3187 .BufferStartingAddress
= anv_address_add(buffer
->address
, offset
),
3190 .BufferSize
= buffer
->size
- offset
3192 .EndAddress
= anv_address_add(buffer
->address
, buffer
->size
- 1),
3196 #if GEN_GEN >= 8 && GEN_GEN <= 9
3197 genX(cmd_buffer_set_binding_for_gen8_vb_flush
)(cmd_buffer
, vb
,
3198 state
.BufferStartingAddress
,
3202 GENX(VERTEX_BUFFER_STATE_pack
)(&cmd_buffer
->batch
, &p
[1 + i
* 4], &state
);
3207 cmd_buffer
->state
.gfx
.vb_dirty
&= ~vb_emit
;
3210 if (cmd_buffer
->state
.gfx
.dirty
& ANV_CMD_DIRTY_XFB_ENABLE
) {
3211 /* We don't need any per-buffer dirty tracking because you're not
3212 * allowed to bind different XFB buffers while XFB is enabled.
3214 for (unsigned idx
= 0; idx
< MAX_XFB_BUFFERS
; idx
++) {
3215 struct anv_xfb_binding
*xfb
= &cmd_buffer
->state
.xfb_bindings
[idx
];
3216 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DSTATE_SO_BUFFER
), sob
) {
3218 sob
.SOBufferIndex
= idx
;
3220 sob
._3DCommandOpcode
= 0;
3221 sob
._3DCommandSubOpcode
= SO_BUFFER_INDEX_0_CMD
+ idx
;
3224 if (cmd_buffer
->state
.xfb_enabled
&& xfb
->buffer
&& xfb
->size
!= 0) {
3225 sob
.SOBufferEnable
= true;
3226 sob
.MOCS
= cmd_buffer
->device
->isl_dev
.mocs
.internal
,
3227 sob
.StreamOffsetWriteEnable
= false;
3228 sob
.SurfaceBaseAddress
= anv_address_add(xfb
->buffer
->address
,
3230 /* Size is in DWords - 1 */
3231 sob
.SurfaceSize
= xfb
->size
/ 4 - 1;
3236 /* CNL and later require a CS stall after 3DSTATE_SO_BUFFER */
3238 cmd_buffer
->state
.pending_pipe_bits
|= ANV_PIPE_CS_STALL_BIT
;
3242 if (cmd_buffer
->state
.gfx
.dirty
& ANV_CMD_DIRTY_PIPELINE
) {
3243 anv_batch_emit_batch(&cmd_buffer
->batch
, &pipeline
->batch
);
3245 /* If the pipeline changed, we may need to re-allocate push constant
3248 cmd_buffer_alloc_push_constants(cmd_buffer
);
3252 if (cmd_buffer
->state
.descriptors_dirty
& VK_SHADER_STAGE_VERTEX_BIT
||
3253 cmd_buffer
->state
.push_constants_dirty
& VK_SHADER_STAGE_VERTEX_BIT
) {
3254 /* From the IVB PRM Vol. 2, Part 1, Section 3.2.1:
3256 * "A PIPE_CONTROL with Post-Sync Operation set to 1h and a depth
3257 * stall needs to be sent just prior to any 3DSTATE_VS,
3258 * 3DSTATE_URB_VS, 3DSTATE_CONSTANT_VS,
3259 * 3DSTATE_BINDING_TABLE_POINTER_VS,
3260 * 3DSTATE_SAMPLER_STATE_POINTER_VS command. Only one
3261 * PIPE_CONTROL needs to be sent before any combination of VS
3262 * associated 3DSTATE."
3264 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
3265 pc
.DepthStallEnable
= true;
3266 pc
.PostSyncOperation
= WriteImmediateData
;
3268 (struct anv_address
) { cmd_buffer
->device
->workaround_bo
, 0 };
3273 /* Render targets live in the same binding table as fragment descriptors */
3274 if (cmd_buffer
->state
.gfx
.dirty
& ANV_CMD_DIRTY_RENDER_TARGETS
)
3275 cmd_buffer
->state
.descriptors_dirty
|= VK_SHADER_STAGE_FRAGMENT_BIT
;
3277 /* We emit the binding tables and sampler tables first, then emit push
3278 * constants and then finally emit binding table and sampler table
3279 * pointers. It has to happen in this order, since emitting the binding
3280 * tables may change the push constants (in case of storage images). After
3281 * emitting push constants, on SKL+ we have to emit the corresponding
3282 * 3DSTATE_BINDING_TABLE_POINTER_* for the push constants to take effect.
3285 if (cmd_buffer
->state
.descriptors_dirty
)
3286 dirty
= flush_descriptor_sets(cmd_buffer
, pipeline
);
3288 if (dirty
|| cmd_buffer
->state
.push_constants_dirty
) {
3289 /* Because we're pushing UBOs, we have to push whenever either
3290 * descriptors or push constants is dirty.
3292 dirty
|= cmd_buffer
->state
.push_constants_dirty
;
3293 dirty
&= ANV_STAGE_MASK
& VK_SHADER_STAGE_ALL_GRAPHICS
;
3294 cmd_buffer_flush_push_constants(cmd_buffer
, dirty
);
3298 cmd_buffer_emit_descriptor_pointers(cmd_buffer
, dirty
);
3300 if (cmd_buffer
->state
.gfx
.dirty
& ANV_CMD_DIRTY_DYNAMIC_VIEWPORT
)
3301 gen8_cmd_buffer_emit_viewport(cmd_buffer
);
3303 if (cmd_buffer
->state
.gfx
.dirty
& (ANV_CMD_DIRTY_DYNAMIC_VIEWPORT
|
3304 ANV_CMD_DIRTY_PIPELINE
)) {
3305 gen8_cmd_buffer_emit_depth_viewport(cmd_buffer
,
3306 pipeline
->depth_clamp_enable
);
3309 if (cmd_buffer
->state
.gfx
.dirty
& (ANV_CMD_DIRTY_DYNAMIC_SCISSOR
|
3310 ANV_CMD_DIRTY_RENDER_TARGETS
))
3311 gen7_cmd_buffer_emit_scissor(cmd_buffer
);
3313 genX(cmd_buffer_flush_dynamic_state
)(cmd_buffer
);
3317 emit_vertex_bo(struct anv_cmd_buffer
*cmd_buffer
,
3318 struct anv_address addr
,
3319 uint32_t size
, uint32_t index
)
3321 uint32_t *p
= anv_batch_emitn(&cmd_buffer
->batch
, 5,
3322 GENX(3DSTATE_VERTEX_BUFFERS
));
3324 GENX(VERTEX_BUFFER_STATE_pack
)(&cmd_buffer
->batch
, p
+ 1,
3325 &(struct GENX(VERTEX_BUFFER_STATE
)) {
3326 .VertexBufferIndex
= index
,
3327 .AddressModifyEnable
= true,
3329 .MOCS
= addr
.bo
? anv_mocs_for_bo(cmd_buffer
->device
, addr
.bo
) : 0,
3330 .NullVertexBuffer
= size
== 0,
3332 .BufferStartingAddress
= addr
,
3335 .BufferStartingAddress
= addr
,
3336 .EndAddress
= anv_address_add(addr
, size
),
3340 genX(cmd_buffer_set_binding_for_gen8_vb_flush
)(cmd_buffer
,
3345 emit_base_vertex_instance_bo(struct anv_cmd_buffer
*cmd_buffer
,
3346 struct anv_address addr
)
3348 emit_vertex_bo(cmd_buffer
, addr
, addr
.bo
? 8 : 0, ANV_SVGS_VB_INDEX
);
3352 emit_base_vertex_instance(struct anv_cmd_buffer
*cmd_buffer
,
3353 uint32_t base_vertex
, uint32_t base_instance
)
3355 if (base_vertex
== 0 && base_instance
== 0) {
3356 emit_base_vertex_instance_bo(cmd_buffer
, ANV_NULL_ADDRESS
);
3358 struct anv_state id_state
=
3359 anv_cmd_buffer_alloc_dynamic_state(cmd_buffer
, 8, 4);
3361 ((uint32_t *)id_state
.map
)[0] = base_vertex
;
3362 ((uint32_t *)id_state
.map
)[1] = base_instance
;
3364 struct anv_address addr
= {
3365 .bo
= cmd_buffer
->device
->dynamic_state_pool
.block_pool
.bo
,
3366 .offset
= id_state
.offset
,
3369 emit_base_vertex_instance_bo(cmd_buffer
, addr
);
3374 emit_draw_index(struct anv_cmd_buffer
*cmd_buffer
, uint32_t draw_index
)
3376 struct anv_state state
=
3377 anv_cmd_buffer_alloc_dynamic_state(cmd_buffer
, 4, 4);
3379 ((uint32_t *)state
.map
)[0] = draw_index
;
3381 struct anv_address addr
= {
3382 .bo
= cmd_buffer
->device
->dynamic_state_pool
.block_pool
.bo
,
3383 .offset
= state
.offset
,
3386 emit_vertex_bo(cmd_buffer
, addr
, 4, ANV_DRAWID_VB_INDEX
);
3390 update_dirty_vbs_for_gen8_vb_flush(struct anv_cmd_buffer
*cmd_buffer
,
3391 uint32_t access_type
)
3393 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.gfx
.base
.pipeline
;
3394 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
3396 uint64_t vb_used
= pipeline
->vb_used
;
3397 if (vs_prog_data
->uses_firstvertex
||
3398 vs_prog_data
->uses_baseinstance
)
3399 vb_used
|= 1ull << ANV_SVGS_VB_INDEX
;
3400 if (vs_prog_data
->uses_drawid
)
3401 vb_used
|= 1ull << ANV_DRAWID_VB_INDEX
;
3403 genX(cmd_buffer_update_dirty_vbs_for_gen8_vb_flush
)(cmd_buffer
,
3404 access_type
== RANDOM
,
3409 VkCommandBuffer commandBuffer
,
3410 uint32_t vertexCount
,
3411 uint32_t instanceCount
,
3412 uint32_t firstVertex
,
3413 uint32_t firstInstance
)
3415 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3416 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.gfx
.base
.pipeline
;
3417 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
3419 if (anv_batch_has_error(&cmd_buffer
->batch
))
3422 genX(cmd_buffer_flush_state
)(cmd_buffer
);
3424 if (cmd_buffer
->state
.conditional_render_enabled
)
3425 genX(cmd_emit_conditional_render_predicate
)(cmd_buffer
);
3427 if (vs_prog_data
->uses_firstvertex
||
3428 vs_prog_data
->uses_baseinstance
)
3429 emit_base_vertex_instance(cmd_buffer
, firstVertex
, firstInstance
);
3430 if (vs_prog_data
->uses_drawid
)
3431 emit_draw_index(cmd_buffer
, 0);
3433 /* Emitting draw index or vertex index BOs may result in needing
3434 * additional VF cache flushes.
3436 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
3438 /* Our implementation of VK_KHR_multiview uses instancing to draw the
3439 * different views. We need to multiply instanceCount by the view count.
3441 instanceCount
*= anv_subpass_view_count(cmd_buffer
->state
.subpass
);
3443 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DPRIMITIVE
), prim
) {
3444 prim
.PredicateEnable
= cmd_buffer
->state
.conditional_render_enabled
;
3445 prim
.VertexAccessType
= SEQUENTIAL
;
3446 prim
.PrimitiveTopologyType
= pipeline
->topology
;
3447 prim
.VertexCountPerInstance
= vertexCount
;
3448 prim
.StartVertexLocation
= firstVertex
;
3449 prim
.InstanceCount
= instanceCount
;
3450 prim
.StartInstanceLocation
= firstInstance
;
3451 prim
.BaseVertexLocation
= 0;
3454 update_dirty_vbs_for_gen8_vb_flush(cmd_buffer
, SEQUENTIAL
);
3457 void genX(CmdDrawIndexed
)(
3458 VkCommandBuffer commandBuffer
,
3459 uint32_t indexCount
,
3460 uint32_t instanceCount
,
3461 uint32_t firstIndex
,
3462 int32_t vertexOffset
,
3463 uint32_t firstInstance
)
3465 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3466 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.gfx
.base
.pipeline
;
3467 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
3469 if (anv_batch_has_error(&cmd_buffer
->batch
))
3472 genX(cmd_buffer_flush_state
)(cmd_buffer
);
3474 if (cmd_buffer
->state
.conditional_render_enabled
)
3475 genX(cmd_emit_conditional_render_predicate
)(cmd_buffer
);
3477 if (vs_prog_data
->uses_firstvertex
||
3478 vs_prog_data
->uses_baseinstance
)
3479 emit_base_vertex_instance(cmd_buffer
, vertexOffset
, firstInstance
);
3480 if (vs_prog_data
->uses_drawid
)
3481 emit_draw_index(cmd_buffer
, 0);
3483 /* Emitting draw index or vertex index BOs may result in needing
3484 * additional VF cache flushes.
3486 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
3488 /* Our implementation of VK_KHR_multiview uses instancing to draw the
3489 * different views. We need to multiply instanceCount by the view count.
3491 instanceCount
*= anv_subpass_view_count(cmd_buffer
->state
.subpass
);
3493 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DPRIMITIVE
), prim
) {
3494 prim
.PredicateEnable
= cmd_buffer
->state
.conditional_render_enabled
;
3495 prim
.VertexAccessType
= RANDOM
;
3496 prim
.PrimitiveTopologyType
= pipeline
->topology
;
3497 prim
.VertexCountPerInstance
= indexCount
;
3498 prim
.StartVertexLocation
= firstIndex
;
3499 prim
.InstanceCount
= instanceCount
;
3500 prim
.StartInstanceLocation
= firstInstance
;
3501 prim
.BaseVertexLocation
= vertexOffset
;
3504 update_dirty_vbs_for_gen8_vb_flush(cmd_buffer
, RANDOM
);
3507 /* Auto-Draw / Indirect Registers */
3508 #define GEN7_3DPRIM_END_OFFSET 0x2420
3509 #define GEN7_3DPRIM_START_VERTEX 0x2430
3510 #define GEN7_3DPRIM_VERTEX_COUNT 0x2434
3511 #define GEN7_3DPRIM_INSTANCE_COUNT 0x2438
3512 #define GEN7_3DPRIM_START_INSTANCE 0x243C
3513 #define GEN7_3DPRIM_BASE_VERTEX 0x2440
3515 void genX(CmdDrawIndirectByteCountEXT
)(
3516 VkCommandBuffer commandBuffer
,
3517 uint32_t instanceCount
,
3518 uint32_t firstInstance
,
3519 VkBuffer counterBuffer
,
3520 VkDeviceSize counterBufferOffset
,
3521 uint32_t counterOffset
,
3522 uint32_t vertexStride
)
3524 #if GEN_IS_HASWELL || GEN_GEN >= 8
3525 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3526 ANV_FROM_HANDLE(anv_buffer
, counter_buffer
, counterBuffer
);
3527 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.gfx
.base
.pipeline
;
3528 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
3530 /* firstVertex is always zero for this draw function */
3531 const uint32_t firstVertex
= 0;
3533 if (anv_batch_has_error(&cmd_buffer
->batch
))
3536 genX(cmd_buffer_flush_state
)(cmd_buffer
);
3538 if (vs_prog_data
->uses_firstvertex
||
3539 vs_prog_data
->uses_baseinstance
)
3540 emit_base_vertex_instance(cmd_buffer
, firstVertex
, firstInstance
);
3541 if (vs_prog_data
->uses_drawid
)
3542 emit_draw_index(cmd_buffer
, 0);
3544 /* Emitting draw index or vertex index BOs may result in needing
3545 * additional VF cache flushes.
3547 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
3549 /* Our implementation of VK_KHR_multiview uses instancing to draw the
3550 * different views. We need to multiply instanceCount by the view count.
3552 instanceCount
*= anv_subpass_view_count(cmd_buffer
->state
.subpass
);
3554 struct gen_mi_builder b
;
3555 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
3556 struct gen_mi_value count
=
3557 gen_mi_mem32(anv_address_add(counter_buffer
->address
,
3558 counterBufferOffset
));
3560 count
= gen_mi_isub(&b
, count
, gen_mi_imm(counterOffset
));
3561 count
= gen_mi_udiv32_imm(&b
, count
, vertexStride
);
3562 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_VERTEX_COUNT
), count
);
3564 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_START_VERTEX
),
3565 gen_mi_imm(firstVertex
));
3566 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_INSTANCE_COUNT
),
3567 gen_mi_imm(instanceCount
));
3568 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_START_INSTANCE
),
3569 gen_mi_imm(firstInstance
));
3570 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_BASE_VERTEX
), gen_mi_imm(0));
3572 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DPRIMITIVE
), prim
) {
3573 prim
.IndirectParameterEnable
= true;
3574 prim
.VertexAccessType
= SEQUENTIAL
;
3575 prim
.PrimitiveTopologyType
= pipeline
->topology
;
3578 update_dirty_vbs_for_gen8_vb_flush(cmd_buffer
, SEQUENTIAL
);
3579 #endif /* GEN_IS_HASWELL || GEN_GEN >= 8 */
3583 load_indirect_parameters(struct anv_cmd_buffer
*cmd_buffer
,
3584 struct anv_address addr
,
3587 struct gen_mi_builder b
;
3588 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
3590 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_VERTEX_COUNT
),
3591 gen_mi_mem32(anv_address_add(addr
, 0)));
3593 struct gen_mi_value instance_count
= gen_mi_mem32(anv_address_add(addr
, 4));
3594 unsigned view_count
= anv_subpass_view_count(cmd_buffer
->state
.subpass
);
3595 if (view_count
> 1) {
3596 #if GEN_IS_HASWELL || GEN_GEN >= 8
3597 instance_count
= gen_mi_imul_imm(&b
, instance_count
, view_count
);
3599 anv_finishme("Multiview + indirect draw requires MI_MATH; "
3600 "MI_MATH is not supported on Ivy Bridge");
3603 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_INSTANCE_COUNT
), instance_count
);
3605 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_START_VERTEX
),
3606 gen_mi_mem32(anv_address_add(addr
, 8)));
3609 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_BASE_VERTEX
),
3610 gen_mi_mem32(anv_address_add(addr
, 12)));
3611 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_START_INSTANCE
),
3612 gen_mi_mem32(anv_address_add(addr
, 16)));
3614 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_START_INSTANCE
),
3615 gen_mi_mem32(anv_address_add(addr
, 12)));
3616 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_BASE_VERTEX
), gen_mi_imm(0));
3620 void genX(CmdDrawIndirect
)(
3621 VkCommandBuffer commandBuffer
,
3623 VkDeviceSize offset
,
3627 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3628 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3629 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.gfx
.base
.pipeline
;
3630 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
3632 if (anv_batch_has_error(&cmd_buffer
->batch
))
3635 genX(cmd_buffer_flush_state
)(cmd_buffer
);
3637 if (cmd_buffer
->state
.conditional_render_enabled
)
3638 genX(cmd_emit_conditional_render_predicate
)(cmd_buffer
);
3640 for (uint32_t i
= 0; i
< drawCount
; i
++) {
3641 struct anv_address draw
= anv_address_add(buffer
->address
, offset
);
3643 if (vs_prog_data
->uses_firstvertex
||
3644 vs_prog_data
->uses_baseinstance
)
3645 emit_base_vertex_instance_bo(cmd_buffer
, anv_address_add(draw
, 8));
3646 if (vs_prog_data
->uses_drawid
)
3647 emit_draw_index(cmd_buffer
, i
);
3649 /* Emitting draw index or vertex index BOs may result in needing
3650 * additional VF cache flushes.
3652 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
3654 load_indirect_parameters(cmd_buffer
, draw
, false);
3656 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DPRIMITIVE
), prim
) {
3657 prim
.IndirectParameterEnable
= true;
3658 prim
.PredicateEnable
= cmd_buffer
->state
.conditional_render_enabled
;
3659 prim
.VertexAccessType
= SEQUENTIAL
;
3660 prim
.PrimitiveTopologyType
= pipeline
->topology
;
3663 update_dirty_vbs_for_gen8_vb_flush(cmd_buffer
, SEQUENTIAL
);
3669 void genX(CmdDrawIndexedIndirect
)(
3670 VkCommandBuffer commandBuffer
,
3672 VkDeviceSize offset
,
3676 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3677 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3678 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.gfx
.base
.pipeline
;
3679 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
3681 if (anv_batch_has_error(&cmd_buffer
->batch
))
3684 genX(cmd_buffer_flush_state
)(cmd_buffer
);
3686 if (cmd_buffer
->state
.conditional_render_enabled
)
3687 genX(cmd_emit_conditional_render_predicate
)(cmd_buffer
);
3689 for (uint32_t i
= 0; i
< drawCount
; i
++) {
3690 struct anv_address draw
= anv_address_add(buffer
->address
, offset
);
3692 /* TODO: We need to stomp base vertex to 0 somehow */
3693 if (vs_prog_data
->uses_firstvertex
||
3694 vs_prog_data
->uses_baseinstance
)
3695 emit_base_vertex_instance_bo(cmd_buffer
, anv_address_add(draw
, 12));
3696 if (vs_prog_data
->uses_drawid
)
3697 emit_draw_index(cmd_buffer
, i
);
3699 /* Emitting draw index or vertex index BOs may result in needing
3700 * additional VF cache flushes.
3702 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
3704 load_indirect_parameters(cmd_buffer
, draw
, true);
3706 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DPRIMITIVE
), prim
) {
3707 prim
.IndirectParameterEnable
= true;
3708 prim
.PredicateEnable
= cmd_buffer
->state
.conditional_render_enabled
;
3709 prim
.VertexAccessType
= RANDOM
;
3710 prim
.PrimitiveTopologyType
= pipeline
->topology
;
3713 update_dirty_vbs_for_gen8_vb_flush(cmd_buffer
, RANDOM
);
3719 #define TMP_DRAW_COUNT_REG 0x2670 /* MI_ALU_REG14 */
3722 prepare_for_draw_count_predicate(struct anv_cmd_buffer
*cmd_buffer
,
3723 struct anv_address count_address
,
3724 const bool conditional_render_enabled
)
3726 struct gen_mi_builder b
;
3727 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
3729 if (conditional_render_enabled
) {
3730 #if GEN_GEN >= 8 || GEN_IS_HASWELL
3731 gen_mi_store(&b
, gen_mi_reg64(TMP_DRAW_COUNT_REG
),
3732 gen_mi_mem32(count_address
));
3735 /* Upload the current draw count from the draw parameters buffer to
3736 * MI_PREDICATE_SRC0.
3738 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC0
),
3739 gen_mi_mem32(count_address
));
3741 gen_mi_store(&b
, gen_mi_reg32(MI_PREDICATE_SRC1
+ 4), gen_mi_imm(0));
3746 emit_draw_count_predicate(struct anv_cmd_buffer
*cmd_buffer
,
3747 uint32_t draw_index
)
3749 struct gen_mi_builder b
;
3750 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
3752 /* Upload the index of the current primitive to MI_PREDICATE_SRC1. */
3753 gen_mi_store(&b
, gen_mi_reg32(MI_PREDICATE_SRC1
), gen_mi_imm(draw_index
));
3755 if (draw_index
== 0) {
3756 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_PREDICATE
), mip
) {
3757 mip
.LoadOperation
= LOAD_LOADINV
;
3758 mip
.CombineOperation
= COMBINE_SET
;
3759 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
3762 /* While draw_index < draw_count the predicate's result will be
3763 * (draw_index == draw_count) ^ TRUE = TRUE
3764 * When draw_index == draw_count the result is
3765 * (TRUE) ^ TRUE = FALSE
3766 * After this all results will be:
3767 * (FALSE) ^ FALSE = FALSE
3769 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_PREDICATE
), mip
) {
3770 mip
.LoadOperation
= LOAD_LOAD
;
3771 mip
.CombineOperation
= COMBINE_XOR
;
3772 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
3777 #if GEN_GEN >= 8 || GEN_IS_HASWELL
3779 emit_draw_count_predicate_with_conditional_render(
3780 struct anv_cmd_buffer
*cmd_buffer
,
3781 uint32_t draw_index
)
3783 struct gen_mi_builder b
;
3784 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
3786 struct gen_mi_value pred
= gen_mi_ult(&b
, gen_mi_imm(draw_index
),
3787 gen_mi_reg64(TMP_DRAW_COUNT_REG
));
3788 pred
= gen_mi_iand(&b
, pred
, gen_mi_reg64(ANV_PREDICATE_RESULT_REG
));
3791 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_RESULT
), pred
);
3793 /* MI_PREDICATE_RESULT is not whitelisted in i915 command parser
3794 * so we emit MI_PREDICATE to set it.
3797 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC0
), pred
);
3798 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC1
), gen_mi_imm(0));
3800 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_PREDICATE
), mip
) {
3801 mip
.LoadOperation
= LOAD_LOADINV
;
3802 mip
.CombineOperation
= COMBINE_SET
;
3803 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
3809 void genX(CmdDrawIndirectCount
)(
3810 VkCommandBuffer commandBuffer
,
3812 VkDeviceSize offset
,
3813 VkBuffer _countBuffer
,
3814 VkDeviceSize countBufferOffset
,
3815 uint32_t maxDrawCount
,
3818 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3819 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3820 ANV_FROM_HANDLE(anv_buffer
, count_buffer
, _countBuffer
);
3821 struct anv_cmd_state
*cmd_state
= &cmd_buffer
->state
;
3822 struct anv_pipeline
*pipeline
= cmd_state
->gfx
.base
.pipeline
;
3823 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
3825 if (anv_batch_has_error(&cmd_buffer
->batch
))
3828 genX(cmd_buffer_flush_state
)(cmd_buffer
);
3830 struct anv_address count_address
=
3831 anv_address_add(count_buffer
->address
, countBufferOffset
);
3833 prepare_for_draw_count_predicate(cmd_buffer
, count_address
,
3834 cmd_state
->conditional_render_enabled
);
3836 for (uint32_t i
= 0; i
< maxDrawCount
; i
++) {
3837 struct anv_address draw
= anv_address_add(buffer
->address
, offset
);
3839 #if GEN_GEN >= 8 || GEN_IS_HASWELL
3840 if (cmd_state
->conditional_render_enabled
) {
3841 emit_draw_count_predicate_with_conditional_render(cmd_buffer
, i
);
3843 emit_draw_count_predicate(cmd_buffer
, i
);
3846 emit_draw_count_predicate(cmd_buffer
, i
);
3849 if (vs_prog_data
->uses_firstvertex
||
3850 vs_prog_data
->uses_baseinstance
)
3851 emit_base_vertex_instance_bo(cmd_buffer
, anv_address_add(draw
, 8));
3852 if (vs_prog_data
->uses_drawid
)
3853 emit_draw_index(cmd_buffer
, i
);
3855 /* Emitting draw index or vertex index BOs may result in needing
3856 * additional VF cache flushes.
3858 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
3860 load_indirect_parameters(cmd_buffer
, draw
, false);
3862 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DPRIMITIVE
), prim
) {
3863 prim
.IndirectParameterEnable
= true;
3864 prim
.PredicateEnable
= true;
3865 prim
.VertexAccessType
= SEQUENTIAL
;
3866 prim
.PrimitiveTopologyType
= pipeline
->topology
;
3869 update_dirty_vbs_for_gen8_vb_flush(cmd_buffer
, SEQUENTIAL
);
3875 void genX(CmdDrawIndexedIndirectCount
)(
3876 VkCommandBuffer commandBuffer
,
3878 VkDeviceSize offset
,
3879 VkBuffer _countBuffer
,
3880 VkDeviceSize countBufferOffset
,
3881 uint32_t maxDrawCount
,
3884 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3885 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3886 ANV_FROM_HANDLE(anv_buffer
, count_buffer
, _countBuffer
);
3887 struct anv_cmd_state
*cmd_state
= &cmd_buffer
->state
;
3888 struct anv_pipeline
*pipeline
= cmd_state
->gfx
.base
.pipeline
;
3889 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
3891 if (anv_batch_has_error(&cmd_buffer
->batch
))
3894 genX(cmd_buffer_flush_state
)(cmd_buffer
);
3896 struct anv_address count_address
=
3897 anv_address_add(count_buffer
->address
, countBufferOffset
);
3899 prepare_for_draw_count_predicate(cmd_buffer
, count_address
,
3900 cmd_state
->conditional_render_enabled
);
3902 for (uint32_t i
= 0; i
< maxDrawCount
; i
++) {
3903 struct anv_address draw
= anv_address_add(buffer
->address
, offset
);
3905 #if GEN_GEN >= 8 || GEN_IS_HASWELL
3906 if (cmd_state
->conditional_render_enabled
) {
3907 emit_draw_count_predicate_with_conditional_render(cmd_buffer
, i
);
3909 emit_draw_count_predicate(cmd_buffer
, i
);
3912 emit_draw_count_predicate(cmd_buffer
, i
);
3915 /* TODO: We need to stomp base vertex to 0 somehow */
3916 if (vs_prog_data
->uses_firstvertex
||
3917 vs_prog_data
->uses_baseinstance
)
3918 emit_base_vertex_instance_bo(cmd_buffer
, anv_address_add(draw
, 12));
3919 if (vs_prog_data
->uses_drawid
)
3920 emit_draw_index(cmd_buffer
, i
);
3922 /* Emitting draw index or vertex index BOs may result in needing
3923 * additional VF cache flushes.
3925 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
3927 load_indirect_parameters(cmd_buffer
, draw
, true);
3929 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DPRIMITIVE
), prim
) {
3930 prim
.IndirectParameterEnable
= true;
3931 prim
.PredicateEnable
= true;
3932 prim
.VertexAccessType
= RANDOM
;
3933 prim
.PrimitiveTopologyType
= pipeline
->topology
;
3936 update_dirty_vbs_for_gen8_vb_flush(cmd_buffer
, RANDOM
);
3942 void genX(CmdBeginTransformFeedbackEXT
)(
3943 VkCommandBuffer commandBuffer
,
3944 uint32_t firstCounterBuffer
,
3945 uint32_t counterBufferCount
,
3946 const VkBuffer
* pCounterBuffers
,
3947 const VkDeviceSize
* pCounterBufferOffsets
)
3949 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3951 assert(firstCounterBuffer
< MAX_XFB_BUFFERS
);
3952 assert(counterBufferCount
<= MAX_XFB_BUFFERS
);
3953 assert(firstCounterBuffer
+ counterBufferCount
<= MAX_XFB_BUFFERS
);
3955 /* From the SKL PRM Vol. 2c, SO_WRITE_OFFSET:
3957 * "Ssoftware must ensure that no HW stream output operations can be in
3958 * process or otherwise pending at the point that the MI_LOAD/STORE
3959 * commands are processed. This will likely require a pipeline flush."
3961 cmd_buffer
->state
.pending_pipe_bits
|= ANV_PIPE_CS_STALL_BIT
;
3962 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
3964 for (uint32_t idx
= 0; idx
< MAX_XFB_BUFFERS
; idx
++) {
3965 /* If we have a counter buffer, this is a resume so we need to load the
3966 * value into the streamout offset register. Otherwise, this is a begin
3967 * and we need to reset it to zero.
3969 if (pCounterBuffers
&&
3970 idx
>= firstCounterBuffer
&&
3971 idx
- firstCounterBuffer
< counterBufferCount
&&
3972 pCounterBuffers
[idx
- firstCounterBuffer
] != VK_NULL_HANDLE
) {
3973 uint32_t cb_idx
= idx
- firstCounterBuffer
;
3974 ANV_FROM_HANDLE(anv_buffer
, counter_buffer
, pCounterBuffers
[cb_idx
]);
3975 uint64_t offset
= pCounterBufferOffsets
?
3976 pCounterBufferOffsets
[cb_idx
] : 0;
3978 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
3979 lrm
.RegisterAddress
= GENX(SO_WRITE_OFFSET0_num
) + idx
* 4;
3980 lrm
.MemoryAddress
= anv_address_add(counter_buffer
->address
,
3984 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_LOAD_REGISTER_IMM
), lri
) {
3985 lri
.RegisterOffset
= GENX(SO_WRITE_OFFSET0_num
) + idx
* 4;
3991 cmd_buffer
->state
.xfb_enabled
= true;
3992 cmd_buffer
->state
.gfx
.dirty
|= ANV_CMD_DIRTY_XFB_ENABLE
;
3995 void genX(CmdEndTransformFeedbackEXT
)(
3996 VkCommandBuffer commandBuffer
,
3997 uint32_t firstCounterBuffer
,
3998 uint32_t counterBufferCount
,
3999 const VkBuffer
* pCounterBuffers
,
4000 const VkDeviceSize
* pCounterBufferOffsets
)
4002 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
4004 assert(firstCounterBuffer
< MAX_XFB_BUFFERS
);
4005 assert(counterBufferCount
<= MAX_XFB_BUFFERS
);
4006 assert(firstCounterBuffer
+ counterBufferCount
<= MAX_XFB_BUFFERS
);
4008 /* From the SKL PRM Vol. 2c, SO_WRITE_OFFSET:
4010 * "Ssoftware must ensure that no HW stream output operations can be in
4011 * process or otherwise pending at the point that the MI_LOAD/STORE
4012 * commands are processed. This will likely require a pipeline flush."
4014 cmd_buffer
->state
.pending_pipe_bits
|= ANV_PIPE_CS_STALL_BIT
;
4015 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
4017 for (uint32_t cb_idx
= 0; cb_idx
< counterBufferCount
; cb_idx
++) {
4018 unsigned idx
= firstCounterBuffer
+ cb_idx
;
4020 /* If we have a counter buffer, this is a resume so we need to load the
4021 * value into the streamout offset register. Otherwise, this is a begin
4022 * and we need to reset it to zero.
4024 if (pCounterBuffers
&&
4025 cb_idx
< counterBufferCount
&&
4026 pCounterBuffers
[cb_idx
] != VK_NULL_HANDLE
) {
4027 ANV_FROM_HANDLE(anv_buffer
, counter_buffer
, pCounterBuffers
[cb_idx
]);
4028 uint64_t offset
= pCounterBufferOffsets
?
4029 pCounterBufferOffsets
[cb_idx
] : 0;
4031 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_STORE_REGISTER_MEM
), srm
) {
4032 srm
.MemoryAddress
= anv_address_add(counter_buffer
->address
,
4034 srm
.RegisterAddress
= GENX(SO_WRITE_OFFSET0_num
) + idx
* 4;
4039 cmd_buffer
->state
.xfb_enabled
= false;
4040 cmd_buffer
->state
.gfx
.dirty
|= ANV_CMD_DIRTY_XFB_ENABLE
;
4044 genX(cmd_buffer_flush_compute_state
)(struct anv_cmd_buffer
*cmd_buffer
)
4046 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.compute
.base
.pipeline
;
4048 assert(pipeline
->active_stages
== VK_SHADER_STAGE_COMPUTE_BIT
);
4050 genX(cmd_buffer_config_l3
)(cmd_buffer
, pipeline
->l3_config
);
4052 genX(flush_pipeline_select_gpgpu
)(cmd_buffer
);
4054 if (cmd_buffer
->state
.compute
.pipeline_dirty
) {
4055 /* From the Sky Lake PRM Vol 2a, MEDIA_VFE_STATE:
4057 * "A stalling PIPE_CONTROL is required before MEDIA_VFE_STATE unless
4058 * the only bits that are changed are scoreboard related: Scoreboard
4059 * Enable, Scoreboard Type, Scoreboard Mask, Scoreboard * Delta. For
4060 * these scoreboard related states, a MEDIA_STATE_FLUSH is
4063 cmd_buffer
->state
.pending_pipe_bits
|= ANV_PIPE_CS_STALL_BIT
;
4064 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
4066 anv_batch_emit_batch(&cmd_buffer
->batch
, &pipeline
->batch
);
4068 /* The workgroup size of the pipeline affects our push constant layout
4069 * so flag push constants as dirty if we change the pipeline.
4071 cmd_buffer
->state
.push_constants_dirty
|= VK_SHADER_STAGE_COMPUTE_BIT
;
4074 if ((cmd_buffer
->state
.descriptors_dirty
& VK_SHADER_STAGE_COMPUTE_BIT
) ||
4075 cmd_buffer
->state
.compute
.pipeline_dirty
) {
4076 flush_descriptor_sets(cmd_buffer
, pipeline
);
4078 uint32_t iface_desc_data_dw
[GENX(INTERFACE_DESCRIPTOR_DATA_length
)];
4079 struct GENX(INTERFACE_DESCRIPTOR_DATA
) desc
= {
4080 .BindingTablePointer
=
4081 cmd_buffer
->state
.binding_tables
[MESA_SHADER_COMPUTE
].offset
,
4082 .SamplerStatePointer
=
4083 cmd_buffer
->state
.samplers
[MESA_SHADER_COMPUTE
].offset
,
4085 GENX(INTERFACE_DESCRIPTOR_DATA_pack
)(NULL
, iface_desc_data_dw
, &desc
);
4087 struct anv_state state
=
4088 anv_cmd_buffer_merge_dynamic(cmd_buffer
, iface_desc_data_dw
,
4089 pipeline
->interface_descriptor_data
,
4090 GENX(INTERFACE_DESCRIPTOR_DATA_length
),
4093 uint32_t size
= GENX(INTERFACE_DESCRIPTOR_DATA_length
) * sizeof(uint32_t);
4094 anv_batch_emit(&cmd_buffer
->batch
,
4095 GENX(MEDIA_INTERFACE_DESCRIPTOR_LOAD
), mid
) {
4096 mid
.InterfaceDescriptorTotalLength
= size
;
4097 mid
.InterfaceDescriptorDataStartAddress
= state
.offset
;
4101 if (cmd_buffer
->state
.push_constants_dirty
& VK_SHADER_STAGE_COMPUTE_BIT
) {
4102 struct anv_state push_state
=
4103 anv_cmd_buffer_cs_push_constants(cmd_buffer
);
4105 if (push_state
.alloc_size
) {
4106 anv_batch_emit(&cmd_buffer
->batch
, GENX(MEDIA_CURBE_LOAD
), curbe
) {
4107 curbe
.CURBETotalDataLength
= push_state
.alloc_size
;
4108 curbe
.CURBEDataStartAddress
= push_state
.offset
;
4112 cmd_buffer
->state
.push_constants_dirty
&= ~VK_SHADER_STAGE_COMPUTE_BIT
;
4115 cmd_buffer
->state
.compute
.pipeline_dirty
= false;
4117 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
4123 verify_cmd_parser(const struct anv_device
*device
,
4124 int required_version
,
4125 const char *function
)
4127 if (device
->physical
->cmd_parser_version
< required_version
) {
4128 return vk_errorf(device
, device
->physical
,
4129 VK_ERROR_FEATURE_NOT_PRESENT
,
4130 "cmd parser version %d is required for %s",
4131 required_version
, function
);
4140 anv_cmd_buffer_push_base_group_id(struct anv_cmd_buffer
*cmd_buffer
,
4141 uint32_t baseGroupX
,
4142 uint32_t baseGroupY
,
4143 uint32_t baseGroupZ
)
4145 if (anv_batch_has_error(&cmd_buffer
->batch
))
4148 struct anv_push_constants
*push
=
4149 &cmd_buffer
->state
.push_constants
[MESA_SHADER_COMPUTE
];
4150 if (push
->cs
.base_work_group_id
[0] != baseGroupX
||
4151 push
->cs
.base_work_group_id
[1] != baseGroupY
||
4152 push
->cs
.base_work_group_id
[2] != baseGroupZ
) {
4153 push
->cs
.base_work_group_id
[0] = baseGroupX
;
4154 push
->cs
.base_work_group_id
[1] = baseGroupY
;
4155 push
->cs
.base_work_group_id
[2] = baseGroupZ
;
4157 cmd_buffer
->state
.push_constants_dirty
|= VK_SHADER_STAGE_COMPUTE_BIT
;
4161 void genX(CmdDispatch
)(
4162 VkCommandBuffer commandBuffer
,
4167 genX(CmdDispatchBase
)(commandBuffer
, 0, 0, 0, x
, y
, z
);
4170 void genX(CmdDispatchBase
)(
4171 VkCommandBuffer commandBuffer
,
4172 uint32_t baseGroupX
,
4173 uint32_t baseGroupY
,
4174 uint32_t baseGroupZ
,
4175 uint32_t groupCountX
,
4176 uint32_t groupCountY
,
4177 uint32_t groupCountZ
)
4179 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
4180 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.compute
.base
.pipeline
;
4181 const struct brw_cs_prog_data
*prog_data
= get_cs_prog_data(pipeline
);
4183 anv_cmd_buffer_push_base_group_id(cmd_buffer
, baseGroupX
,
4184 baseGroupY
, baseGroupZ
);
4186 if (anv_batch_has_error(&cmd_buffer
->batch
))
4189 if (prog_data
->uses_num_work_groups
) {
4190 struct anv_state state
=
4191 anv_cmd_buffer_alloc_dynamic_state(cmd_buffer
, 12, 4);
4192 uint32_t *sizes
= state
.map
;
4193 sizes
[0] = groupCountX
;
4194 sizes
[1] = groupCountY
;
4195 sizes
[2] = groupCountZ
;
4196 cmd_buffer
->state
.compute
.num_workgroups
= (struct anv_address
) {
4197 .bo
= cmd_buffer
->device
->dynamic_state_pool
.block_pool
.bo
,
4198 .offset
= state
.offset
,
4201 /* The num_workgroups buffer goes in the binding table */
4202 cmd_buffer
->state
.descriptors_dirty
|= VK_SHADER_STAGE_COMPUTE_BIT
;
4205 genX(cmd_buffer_flush_compute_state
)(cmd_buffer
);
4207 if (cmd_buffer
->state
.conditional_render_enabled
)
4208 genX(cmd_emit_conditional_render_predicate
)(cmd_buffer
);
4210 anv_batch_emit(&cmd_buffer
->batch
, GENX(GPGPU_WALKER
), ggw
) {
4211 ggw
.PredicateEnable
= cmd_buffer
->state
.conditional_render_enabled
;
4212 ggw
.SIMDSize
= prog_data
->simd_size
/ 16;
4213 ggw
.ThreadDepthCounterMaximum
= 0;
4214 ggw
.ThreadHeightCounterMaximum
= 0;
4215 ggw
.ThreadWidthCounterMaximum
= prog_data
->threads
- 1;
4216 ggw
.ThreadGroupIDXDimension
= groupCountX
;
4217 ggw
.ThreadGroupIDYDimension
= groupCountY
;
4218 ggw
.ThreadGroupIDZDimension
= groupCountZ
;
4219 ggw
.RightExecutionMask
= pipeline
->cs_right_mask
;
4220 ggw
.BottomExecutionMask
= 0xffffffff;
4223 anv_batch_emit(&cmd_buffer
->batch
, GENX(MEDIA_STATE_FLUSH
), msf
);
4226 #define GPGPU_DISPATCHDIMX 0x2500
4227 #define GPGPU_DISPATCHDIMY 0x2504
4228 #define GPGPU_DISPATCHDIMZ 0x2508
4230 void genX(CmdDispatchIndirect
)(
4231 VkCommandBuffer commandBuffer
,
4233 VkDeviceSize offset
)
4235 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
4236 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
4237 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.compute
.base
.pipeline
;
4238 const struct brw_cs_prog_data
*prog_data
= get_cs_prog_data(pipeline
);
4239 struct anv_address addr
= anv_address_add(buffer
->address
, offset
);
4240 struct anv_batch
*batch
= &cmd_buffer
->batch
;
4242 anv_cmd_buffer_push_base_group_id(cmd_buffer
, 0, 0, 0);
4245 /* Linux 4.4 added command parser version 5 which allows the GPGPU
4246 * indirect dispatch registers to be written.
4248 if (verify_cmd_parser(cmd_buffer
->device
, 5,
4249 "vkCmdDispatchIndirect") != VK_SUCCESS
)
4253 if (prog_data
->uses_num_work_groups
) {
4254 cmd_buffer
->state
.compute
.num_workgroups
= addr
;
4256 /* The num_workgroups buffer goes in the binding table */
4257 cmd_buffer
->state
.descriptors_dirty
|= VK_SHADER_STAGE_COMPUTE_BIT
;
4260 genX(cmd_buffer_flush_compute_state
)(cmd_buffer
);
4262 struct gen_mi_builder b
;
4263 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
4265 struct gen_mi_value size_x
= gen_mi_mem32(anv_address_add(addr
, 0));
4266 struct gen_mi_value size_y
= gen_mi_mem32(anv_address_add(addr
, 4));
4267 struct gen_mi_value size_z
= gen_mi_mem32(anv_address_add(addr
, 8));
4269 gen_mi_store(&b
, gen_mi_reg32(GPGPU_DISPATCHDIMX
), size_x
);
4270 gen_mi_store(&b
, gen_mi_reg32(GPGPU_DISPATCHDIMY
), size_y
);
4271 gen_mi_store(&b
, gen_mi_reg32(GPGPU_DISPATCHDIMZ
), size_z
);
4274 /* predicate = (compute_dispatch_indirect_x_size == 0); */
4275 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC0
), size_x
);
4276 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC1
), gen_mi_imm(0));
4277 anv_batch_emit(batch
, GENX(MI_PREDICATE
), mip
) {
4278 mip
.LoadOperation
= LOAD_LOAD
;
4279 mip
.CombineOperation
= COMBINE_SET
;
4280 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
4283 /* predicate |= (compute_dispatch_indirect_y_size == 0); */
4284 gen_mi_store(&b
, gen_mi_reg32(MI_PREDICATE_SRC0
), size_y
);
4285 anv_batch_emit(batch
, GENX(MI_PREDICATE
), mip
) {
4286 mip
.LoadOperation
= LOAD_LOAD
;
4287 mip
.CombineOperation
= COMBINE_OR
;
4288 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
4291 /* predicate |= (compute_dispatch_indirect_z_size == 0); */
4292 gen_mi_store(&b
, gen_mi_reg32(MI_PREDICATE_SRC0
), size_z
);
4293 anv_batch_emit(batch
, GENX(MI_PREDICATE
), mip
) {
4294 mip
.LoadOperation
= LOAD_LOAD
;
4295 mip
.CombineOperation
= COMBINE_OR
;
4296 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
4299 /* predicate = !predicate; */
4300 anv_batch_emit(batch
, GENX(MI_PREDICATE
), mip
) {
4301 mip
.LoadOperation
= LOAD_LOADINV
;
4302 mip
.CombineOperation
= COMBINE_OR
;
4303 mip
.CompareOperation
= COMPARE_FALSE
;
4307 if (cmd_buffer
->state
.conditional_render_enabled
) {
4308 /* predicate &= !(conditional_rendering_predicate == 0); */
4309 gen_mi_store(&b
, gen_mi_reg32(MI_PREDICATE_SRC0
),
4310 gen_mi_reg32(ANV_PREDICATE_RESULT_REG
));
4311 anv_batch_emit(batch
, GENX(MI_PREDICATE
), mip
) {
4312 mip
.LoadOperation
= LOAD_LOADINV
;
4313 mip
.CombineOperation
= COMBINE_AND
;
4314 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
4319 #else /* GEN_GEN > 7 */
4320 if (cmd_buffer
->state
.conditional_render_enabled
)
4321 genX(cmd_emit_conditional_render_predicate
)(cmd_buffer
);
4324 anv_batch_emit(batch
, GENX(GPGPU_WALKER
), ggw
) {
4325 ggw
.IndirectParameterEnable
= true;
4326 ggw
.PredicateEnable
= GEN_GEN
<= 7 ||
4327 cmd_buffer
->state
.conditional_render_enabled
;
4328 ggw
.SIMDSize
= prog_data
->simd_size
/ 16;
4329 ggw
.ThreadDepthCounterMaximum
= 0;
4330 ggw
.ThreadHeightCounterMaximum
= 0;
4331 ggw
.ThreadWidthCounterMaximum
= prog_data
->threads
- 1;
4332 ggw
.RightExecutionMask
= pipeline
->cs_right_mask
;
4333 ggw
.BottomExecutionMask
= 0xffffffff;
4336 anv_batch_emit(batch
, GENX(MEDIA_STATE_FLUSH
), msf
);
4340 genX(flush_pipeline_select
)(struct anv_cmd_buffer
*cmd_buffer
,
4343 UNUSED
const struct gen_device_info
*devinfo
= &cmd_buffer
->device
->info
;
4345 if (cmd_buffer
->state
.current_pipeline
== pipeline
)
4348 #if GEN_GEN >= 8 && GEN_GEN < 10
4349 /* From the Broadwell PRM, Volume 2a: Instructions, PIPELINE_SELECT:
4351 * Software must clear the COLOR_CALC_STATE Valid field in
4352 * 3DSTATE_CC_STATE_POINTERS command prior to send a PIPELINE_SELECT
4353 * with Pipeline Select set to GPGPU.
4355 * The internal hardware docs recommend the same workaround for Gen9
4358 if (pipeline
== GPGPU
)
4359 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DSTATE_CC_STATE_POINTERS
), t
);
4363 if (pipeline
== _3D
) {
4364 /* There is a mid-object preemption workaround which requires you to
4365 * re-emit MEDIA_VFE_STATE after switching from GPGPU to 3D. However,
4366 * even without preemption, we have issues with geometry flickering when
4367 * GPGPU and 3D are back-to-back and this seems to fix it. We don't
4370 const uint32_t subslices
=
4371 MAX2(cmd_buffer
->device
->physical
->subslice_total
, 1);
4372 anv_batch_emit(&cmd_buffer
->batch
, GENX(MEDIA_VFE_STATE
), vfe
) {
4373 vfe
.MaximumNumberofThreads
=
4374 devinfo
->max_cs_threads
* subslices
- 1;
4375 vfe
.NumberofURBEntries
= 2;
4376 vfe
.URBEntryAllocationSize
= 2;
4379 /* We just emitted a dummy MEDIA_VFE_STATE so now that packet is
4380 * invalid. Set the compute pipeline to dirty to force a re-emit of the
4381 * pipeline in case we get back-to-back dispatch calls with the same
4382 * pipeline and a PIPELINE_SELECT in between.
4384 cmd_buffer
->state
.compute
.pipeline_dirty
= true;
4388 /* From "BXML » GT » MI » vol1a GPU Overview » [Instruction]
4389 * PIPELINE_SELECT [DevBWR+]":
4393 * Software must ensure all the write caches are flushed through a
4394 * stalling PIPE_CONTROL command followed by another PIPE_CONTROL
4395 * command to invalidate read only caches prior to programming
4396 * MI_PIPELINE_SELECT command to change the Pipeline Select Mode.
4398 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
4399 pc
.RenderTargetCacheFlushEnable
= true;
4400 pc
.DepthCacheFlushEnable
= true;
4401 pc
.DCFlushEnable
= true;
4402 pc
.PostSyncOperation
= NoWrite
;
4403 pc
.CommandStreamerStallEnable
= true;
4405 pc
.TileCacheFlushEnable
= true;
4407 /* GEN:BUG:1409600907: "PIPE_CONTROL with Depth Stall Enable bit must be
4408 * set with any PIPE_CONTROL with Depth Flush Enable bit set.
4410 pc
.DepthStallEnable
= true;
4414 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
4415 pc
.TextureCacheInvalidationEnable
= true;
4416 pc
.ConstantCacheInvalidationEnable
= true;
4417 pc
.StateCacheInvalidationEnable
= true;
4418 pc
.InstructionCacheInvalidateEnable
= true;
4419 pc
.PostSyncOperation
= NoWrite
;
4421 pc
.TileCacheFlushEnable
= true;
4425 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPELINE_SELECT
), ps
) {
4429 ps
.PipelineSelection
= pipeline
;
4433 if (devinfo
->is_geminilake
) {
4436 * "This chicken bit works around a hardware issue with barrier logic
4437 * encountered when switching between GPGPU and 3D pipelines. To
4438 * workaround the issue, this mode bit should be set after a pipeline
4442 anv_pack_struct(&scec
, GENX(SLICE_COMMON_ECO_CHICKEN1
),
4444 pipeline
== GPGPU
? GLK_BARRIER_MODE_GPGPU
4445 : GLK_BARRIER_MODE_3D_HULL
,
4446 .GLKBarrierModeMask
= 1);
4447 emit_lri(&cmd_buffer
->batch
, GENX(SLICE_COMMON_ECO_CHICKEN1_num
), scec
);
4451 cmd_buffer
->state
.current_pipeline
= pipeline
;
4455 genX(flush_pipeline_select_3d
)(struct anv_cmd_buffer
*cmd_buffer
)
4457 genX(flush_pipeline_select
)(cmd_buffer
, _3D
);
4461 genX(flush_pipeline_select_gpgpu
)(struct anv_cmd_buffer
*cmd_buffer
)
4463 genX(flush_pipeline_select
)(cmd_buffer
, GPGPU
);
4467 genX(cmd_buffer_emit_gen7_depth_flush
)(struct anv_cmd_buffer
*cmd_buffer
)
4472 /* From the Haswell PRM, documentation for 3DSTATE_DEPTH_BUFFER:
4474 * "Restriction: Prior to changing Depth/Stencil Buffer state (i.e., any
4475 * combination of 3DSTATE_DEPTH_BUFFER, 3DSTATE_CLEAR_PARAMS,
4476 * 3DSTATE_STENCIL_BUFFER, 3DSTATE_HIER_DEPTH_BUFFER) SW must first
4477 * issue a pipelined depth stall (PIPE_CONTROL with Depth Stall bit
4478 * set), followed by a pipelined depth cache flush (PIPE_CONTROL with
4479 * Depth Flush Bit set, followed by another pipelined depth stall
4480 * (PIPE_CONTROL with Depth Stall Bit set), unless SW can otherwise
4481 * guarantee that the pipeline from WM onwards is already flushed (e.g.,
4482 * via a preceding MI_FLUSH)."
4484 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pipe
) {
4485 pipe
.DepthStallEnable
= true;
4487 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pipe
) {
4488 pipe
.DepthCacheFlushEnable
= true;
4490 pipe
.TileCacheFlushEnable
= true;
4493 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pipe
) {
4494 pipe
.DepthStallEnable
= true;
4498 /* From the Skylake PRM, 3DSTATE_VERTEX_BUFFERS:
4500 * "The VF cache needs to be invalidated before binding and then using
4501 * Vertex Buffers that overlap with any previously bound Vertex Buffer
4502 * (at a 64B granularity) since the last invalidation. A VF cache
4503 * invalidate is performed by setting the "VF Cache Invalidation Enable"
4504 * bit in PIPE_CONTROL."
4506 * This is implemented by carefully tracking all vertex and index buffer
4507 * bindings and flushing if the cache ever ends up with a range in the cache
4508 * that would exceed 4 GiB. This is implemented in three parts:
4510 * 1. genX(cmd_buffer_set_binding_for_gen8_vb_flush)() which must be called
4511 * every time a 3DSTATE_VERTEX_BUFFER packet is emitted and informs the
4512 * tracking code of the new binding. If this new binding would cause
4513 * the cache to have a too-large range on the next draw call, a pipeline
4514 * stall and VF cache invalidate are added to pending_pipeline_bits.
4516 * 2. genX(cmd_buffer_apply_pipe_flushes)() resets the cache tracking to
4517 * empty whenever we emit a VF invalidate.
4519 * 3. genX(cmd_buffer_update_dirty_vbs_for_gen8_vb_flush)() must be called
4520 * after every 3DPRIMITIVE and copies the bound range into the dirty
4521 * range for each used buffer. This has to be a separate step because
4522 * we don't always re-bind all buffers and so 1. can't know which
4523 * buffers are actually bound.
4526 genX(cmd_buffer_set_binding_for_gen8_vb_flush
)(struct anv_cmd_buffer
*cmd_buffer
,
4528 struct anv_address vb_address
,
4531 if (GEN_GEN
< 8 || GEN_GEN
> 9 ||
4532 !cmd_buffer
->device
->physical
->use_softpin
)
4535 struct anv_vb_cache_range
*bound
, *dirty
;
4536 if (vb_index
== -1) {
4537 bound
= &cmd_buffer
->state
.gfx
.ib_bound_range
;
4538 dirty
= &cmd_buffer
->state
.gfx
.ib_dirty_range
;
4540 assert(vb_index
>= 0);
4541 assert(vb_index
< ARRAY_SIZE(cmd_buffer
->state
.gfx
.vb_bound_ranges
));
4542 assert(vb_index
< ARRAY_SIZE(cmd_buffer
->state
.gfx
.vb_dirty_ranges
));
4543 bound
= &cmd_buffer
->state
.gfx
.vb_bound_ranges
[vb_index
];
4544 dirty
= &cmd_buffer
->state
.gfx
.vb_dirty_ranges
[vb_index
];
4553 assert(vb_address
.bo
&& (vb_address
.bo
->flags
& EXEC_OBJECT_PINNED
));
4554 bound
->start
= gen_48b_address(anv_address_physical(vb_address
));
4555 bound
->end
= bound
->start
+ vb_size
;
4556 assert(bound
->end
> bound
->start
); /* No overflow */
4558 /* Align everything to a cache line */
4559 bound
->start
&= ~(64ull - 1ull);
4560 bound
->end
= align_u64(bound
->end
, 64);
4562 /* Compute the dirty range */
4563 dirty
->start
= MIN2(dirty
->start
, bound
->start
);
4564 dirty
->end
= MAX2(dirty
->end
, bound
->end
);
4566 /* If our range is larger than 32 bits, we have to flush */
4567 assert(bound
->end
- bound
->start
<= (1ull << 32));
4568 if (dirty
->end
- dirty
->start
> (1ull << 32)) {
4569 cmd_buffer
->state
.pending_pipe_bits
|=
4570 ANV_PIPE_CS_STALL_BIT
| ANV_PIPE_VF_CACHE_INVALIDATE_BIT
;
4575 genX(cmd_buffer_update_dirty_vbs_for_gen8_vb_flush
)(struct anv_cmd_buffer
*cmd_buffer
,
4576 uint32_t access_type
,
4579 if (GEN_GEN
< 8 || GEN_GEN
> 9 ||
4580 !cmd_buffer
->device
->physical
->use_softpin
)
4583 if (access_type
== RANDOM
) {
4584 /* We have an index buffer */
4585 struct anv_vb_cache_range
*bound
= &cmd_buffer
->state
.gfx
.ib_bound_range
;
4586 struct anv_vb_cache_range
*dirty
= &cmd_buffer
->state
.gfx
.ib_dirty_range
;
4588 if (bound
->end
> bound
->start
) {
4589 dirty
->start
= MIN2(dirty
->start
, bound
->start
);
4590 dirty
->end
= MAX2(dirty
->end
, bound
->end
);
4594 uint64_t mask
= vb_used
;
4596 int i
= u_bit_scan64(&mask
);
4598 assert(i
< ARRAY_SIZE(cmd_buffer
->state
.gfx
.vb_bound_ranges
));
4599 assert(i
< ARRAY_SIZE(cmd_buffer
->state
.gfx
.vb_dirty_ranges
));
4601 struct anv_vb_cache_range
*bound
, *dirty
;
4602 bound
= &cmd_buffer
->state
.gfx
.vb_bound_ranges
[i
];
4603 dirty
= &cmd_buffer
->state
.gfx
.vb_dirty_ranges
[i
];
4605 if (bound
->end
> bound
->start
) {
4606 dirty
->start
= MIN2(dirty
->start
, bound
->start
);
4607 dirty
->end
= MAX2(dirty
->end
, bound
->end
);
4613 * Update the pixel hashing modes that determine the balancing of PS threads
4614 * across subslices and slices.
4616 * \param width Width bound of the rendering area (already scaled down if \p
4617 * scale is greater than 1).
4618 * \param height Height bound of the rendering area (already scaled down if \p
4619 * scale is greater than 1).
4620 * \param scale The number of framebuffer samples that could potentially be
4621 * affected by an individual channel of the PS thread. This is
4622 * typically one for single-sampled rendering, but for operations
4623 * like CCS resolves and fast clears a single PS invocation may
4624 * update a huge number of pixels, in which case a finer
4625 * balancing is desirable in order to maximally utilize the
4626 * bandwidth available. UINT_MAX can be used as shorthand for
4627 * "finest hashing mode available".
4630 genX(cmd_buffer_emit_hashing_mode
)(struct anv_cmd_buffer
*cmd_buffer
,
4631 unsigned width
, unsigned height
,
4635 const struct gen_device_info
*devinfo
= &cmd_buffer
->device
->info
;
4636 const unsigned slice_hashing
[] = {
4637 /* Because all Gen9 platforms with more than one slice require
4638 * three-way subslice hashing, a single "normal" 16x16 slice hashing
4639 * block is guaranteed to suffer from substantial imbalance, with one
4640 * subslice receiving twice as much work as the other two in the
4643 * The performance impact of that would be particularly severe when
4644 * three-way hashing is also in use for slice balancing (which is the
4645 * case for all Gen9 GT4 platforms), because one of the slices
4646 * receives one every three 16x16 blocks in either direction, which
4647 * is roughly the periodicity of the underlying subslice imbalance
4648 * pattern ("roughly" because in reality the hardware's
4649 * implementation of three-way hashing doesn't do exact modulo 3
4650 * arithmetic, which somewhat decreases the magnitude of this effect
4651 * in practice). This leads to a systematic subslice imbalance
4652 * within that slice regardless of the size of the primitive. The
4653 * 32x32 hashing mode guarantees that the subslice imbalance within a
4654 * single slice hashing block is minimal, largely eliminating this
4658 /* Finest slice hashing mode available. */
4661 const unsigned subslice_hashing
[] = {
4662 /* 16x16 would provide a slight cache locality benefit especially
4663 * visible in the sampler L1 cache efficiency of low-bandwidth
4664 * non-LLC platforms, but it comes at the cost of greater subslice
4665 * imbalance for primitives of dimensions approximately intermediate
4666 * between 16x4 and 16x16.
4669 /* Finest subslice hashing mode available. */
4672 /* Dimensions of the smallest hashing block of a given hashing mode. If
4673 * the rendering area is smaller than this there can't possibly be any
4674 * benefit from switching to this mode, so we optimize out the
4677 const unsigned min_size
[][2] = {
4681 const unsigned idx
= scale
> 1;
4683 if (cmd_buffer
->state
.current_hash_scale
!= scale
&&
4684 (width
> min_size
[idx
][0] || height
> min_size
[idx
][1])) {
4687 anv_pack_struct(>_mode
, GENX(GT_MODE
),
4688 .SliceHashing
= (devinfo
->num_slices
> 1 ? slice_hashing
[idx
] : 0),
4689 .SliceHashingMask
= (devinfo
->num_slices
> 1 ? -1 : 0),
4690 .SubsliceHashing
= subslice_hashing
[idx
],
4691 .SubsliceHashingMask
= -1);
4693 cmd_buffer
->state
.pending_pipe_bits
|=
4694 ANV_PIPE_CS_STALL_BIT
| ANV_PIPE_STALL_AT_SCOREBOARD_BIT
;
4695 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
4697 emit_lri(&cmd_buffer
->batch
, GENX(GT_MODE_num
), gt_mode
);
4699 cmd_buffer
->state
.current_hash_scale
= scale
;
4705 cmd_buffer_emit_depth_stencil(struct anv_cmd_buffer
*cmd_buffer
)
4707 struct anv_device
*device
= cmd_buffer
->device
;
4708 const struct anv_image_view
*iview
=
4709 anv_cmd_buffer_get_depth_stencil_view(cmd_buffer
);
4710 const struct anv_image
*image
= iview
? iview
->image
: NULL
;
4712 /* FIXME: Width and Height are wrong */
4714 genX(cmd_buffer_emit_gen7_depth_flush
)(cmd_buffer
);
4716 uint32_t *dw
= anv_batch_emit_dwords(&cmd_buffer
->batch
,
4717 device
->isl_dev
.ds
.size
/ 4);
4721 struct isl_depth_stencil_hiz_emit_info info
= { };
4724 info
.view
= &iview
->planes
[0].isl
;
4726 if (image
&& (image
->aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
)) {
4727 uint32_t depth_plane
=
4728 anv_image_aspect_to_plane(image
->aspects
, VK_IMAGE_ASPECT_DEPTH_BIT
);
4729 const struct anv_surface
*surface
= &image
->planes
[depth_plane
].surface
;
4731 info
.depth_surf
= &surface
->isl
;
4733 info
.depth_address
=
4734 anv_batch_emit_reloc(&cmd_buffer
->batch
,
4735 dw
+ device
->isl_dev
.ds
.depth_offset
/ 4,
4736 image
->planes
[depth_plane
].address
.bo
,
4737 image
->planes
[depth_plane
].address
.offset
+
4740 anv_mocs_for_bo(device
, image
->planes
[depth_plane
].address
.bo
);
4743 cmd_buffer
->state
.subpass
->depth_stencil_attachment
->attachment
;
4744 info
.hiz_usage
= cmd_buffer
->state
.attachments
[ds
].aux_usage
;
4745 if (info
.hiz_usage
== ISL_AUX_USAGE_HIZ
) {
4746 info
.hiz_surf
= &image
->planes
[depth_plane
].aux_surface
.isl
;
4749 anv_batch_emit_reloc(&cmd_buffer
->batch
,
4750 dw
+ device
->isl_dev
.ds
.hiz_offset
/ 4,
4751 image
->planes
[depth_plane
].address
.bo
,
4752 image
->planes
[depth_plane
].address
.offset
+
4753 image
->planes
[depth_plane
].aux_surface
.offset
);
4755 info
.depth_clear_value
= ANV_HZ_FC_VAL
;
4759 if (image
&& (image
->aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
)) {
4760 uint32_t stencil_plane
=
4761 anv_image_aspect_to_plane(image
->aspects
, VK_IMAGE_ASPECT_STENCIL_BIT
);
4762 const struct anv_surface
*surface
= &image
->planes
[stencil_plane
].surface
;
4764 info
.stencil_surf
= &surface
->isl
;
4766 info
.stencil_address
=
4767 anv_batch_emit_reloc(&cmd_buffer
->batch
,
4768 dw
+ device
->isl_dev
.ds
.stencil_offset
/ 4,
4769 image
->planes
[stencil_plane
].address
.bo
,
4770 image
->planes
[stencil_plane
].address
.offset
+
4773 anv_mocs_for_bo(device
, image
->planes
[stencil_plane
].address
.bo
);
4776 isl_emit_depth_stencil_hiz_s(&device
->isl_dev
, dw
, &info
);
4778 if (GEN_GEN
>= 12) {
4779 cmd_buffer
->state
.pending_pipe_bits
|= ANV_PIPE_POST_SYNC_BIT
;
4780 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
4782 /* GEN:BUG:1408224581
4784 * Workaround: Gen12LP Astep only An additional pipe control with
4785 * post-sync = store dword operation would be required.( w/a is to
4786 * have an additional pipe control after the stencil state whenever
4787 * the surface state bits of this state is changing).
4789 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
4790 pc
.PostSyncOperation
= WriteImmediateData
;
4792 (struct anv_address
) { cmd_buffer
->device
->workaround_bo
, 0 };
4795 cmd_buffer
->state
.hiz_enabled
= info
.hiz_usage
== ISL_AUX_USAGE_HIZ
;
4799 * This ANDs the view mask of the current subpass with the pending clear
4800 * views in the attachment to get the mask of views active in the subpass
4801 * that still need to be cleared.
4803 static inline uint32_t
4804 get_multiview_subpass_clear_mask(const struct anv_cmd_state
*cmd_state
,
4805 const struct anv_attachment_state
*att_state
)
4807 return cmd_state
->subpass
->view_mask
& att_state
->pending_clear_views
;
4811 do_first_layer_clear(const struct anv_cmd_state
*cmd_state
,
4812 const struct anv_attachment_state
*att_state
)
4814 if (!cmd_state
->subpass
->view_mask
)
4817 uint32_t pending_clear_mask
=
4818 get_multiview_subpass_clear_mask(cmd_state
, att_state
);
4820 return pending_clear_mask
& 1;
4824 current_subpass_is_last_for_attachment(const struct anv_cmd_state
*cmd_state
,
4827 const uint32_t last_subpass_idx
=
4828 cmd_state
->pass
->attachments
[att_idx
].last_subpass_idx
;
4829 const struct anv_subpass
*last_subpass
=
4830 &cmd_state
->pass
->subpasses
[last_subpass_idx
];
4831 return last_subpass
== cmd_state
->subpass
;
4835 cmd_buffer_begin_subpass(struct anv_cmd_buffer
*cmd_buffer
,
4836 uint32_t subpass_id
)
4838 struct anv_cmd_state
*cmd_state
= &cmd_buffer
->state
;
4839 struct anv_subpass
*subpass
= &cmd_state
->pass
->subpasses
[subpass_id
];
4840 cmd_state
->subpass
= subpass
;
4842 cmd_buffer
->state
.gfx
.dirty
|= ANV_CMD_DIRTY_RENDER_TARGETS
;
4844 /* Our implementation of VK_KHR_multiview uses instancing to draw the
4845 * different views. If the client asks for instancing, we need to use the
4846 * Instance Data Step Rate to ensure that we repeat the client's
4847 * per-instance data once for each view. Since this bit is in
4848 * VERTEX_BUFFER_STATE on gen7, we need to dirty vertex buffers at the top
4852 cmd_buffer
->state
.gfx
.vb_dirty
|= ~0;
4854 /* It is possible to start a render pass with an old pipeline. Because the
4855 * render pass and subpass index are both baked into the pipeline, this is
4856 * highly unlikely. In order to do so, it requires that you have a render
4857 * pass with a single subpass and that you use that render pass twice
4858 * back-to-back and use the same pipeline at the start of the second render
4859 * pass as at the end of the first. In order to avoid unpredictable issues
4860 * with this edge case, we just dirty the pipeline at the start of every
4863 cmd_buffer
->state
.gfx
.dirty
|= ANV_CMD_DIRTY_PIPELINE
;
4865 /* Accumulate any subpass flushes that need to happen before the subpass */
4866 cmd_buffer
->state
.pending_pipe_bits
|=
4867 cmd_buffer
->state
.pass
->subpass_flushes
[subpass_id
];
4869 VkRect2D render_area
= cmd_buffer
->state
.render_area
;
4870 struct anv_framebuffer
*fb
= cmd_buffer
->state
.framebuffer
;
4872 bool is_multiview
= subpass
->view_mask
!= 0;
4874 for (uint32_t i
= 0; i
< subpass
->attachment_count
; ++i
) {
4875 const uint32_t a
= subpass
->attachments
[i
].attachment
;
4876 if (a
== VK_ATTACHMENT_UNUSED
)
4879 assert(a
< cmd_state
->pass
->attachment_count
);
4880 struct anv_attachment_state
*att_state
= &cmd_state
->attachments
[a
];
4882 struct anv_image_view
*iview
= cmd_state
->attachments
[a
].image_view
;
4883 const struct anv_image
*image
= iview
->image
;
4885 /* A resolve is necessary before use as an input attachment if the clear
4886 * color or auxiliary buffer usage isn't supported by the sampler.
4888 const bool input_needs_resolve
=
4889 (att_state
->fast_clear
&& !att_state
->clear_color_is_zero_one
) ||
4890 att_state
->input_aux_usage
!= att_state
->aux_usage
;
4892 VkImageLayout target_layout
;
4893 if (iview
->aspect_mask
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
&&
4894 !input_needs_resolve
) {
4895 /* Layout transitions before the final only help to enable sampling
4896 * as an input attachment. If the input attachment supports sampling
4897 * using the auxiliary surface, we can skip such transitions by
4898 * making the target layout one that is CCS-aware.
4900 target_layout
= VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL
;
4902 target_layout
= subpass
->attachments
[i
].layout
;
4905 VkImageLayout target_stencil_layout
=
4906 subpass
->attachments
[i
].stencil_layout
;
4908 uint32_t base_layer
, layer_count
;
4909 if (image
->type
== VK_IMAGE_TYPE_3D
) {
4911 layer_count
= anv_minify(iview
->image
->extent
.depth
,
4912 iview
->planes
[0].isl
.base_level
);
4914 base_layer
= iview
->planes
[0].isl
.base_array_layer
;
4915 layer_count
= fb
->layers
;
4918 if (image
->aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) {
4919 assert(image
->aspects
== VK_IMAGE_ASPECT_COLOR_BIT
);
4920 transition_color_buffer(cmd_buffer
, image
, VK_IMAGE_ASPECT_COLOR_BIT
,
4921 iview
->planes
[0].isl
.base_level
, 1,
4922 base_layer
, layer_count
,
4923 att_state
->current_layout
, target_layout
);
4926 if (image
->aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
) {
4927 transition_depth_buffer(cmd_buffer
, image
,
4928 att_state
->current_layout
, target_layout
);
4929 att_state
->aux_usage
=
4930 anv_layout_to_aux_usage(&cmd_buffer
->device
->info
, image
,
4931 VK_IMAGE_ASPECT_DEPTH_BIT
,
4932 VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT
,
4936 if (image
->aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
) {
4937 transition_stencil_buffer(cmd_buffer
, image
,
4938 iview
->planes
[0].isl
.base_level
, 1,
4939 base_layer
, layer_count
,
4940 att_state
->current_stencil_layout
,
4941 target_stencil_layout
);
4943 att_state
->current_layout
= target_layout
;
4944 att_state
->current_stencil_layout
= target_stencil_layout
;
4946 if (att_state
->pending_clear_aspects
& VK_IMAGE_ASPECT_COLOR_BIT
) {
4947 assert(att_state
->pending_clear_aspects
== VK_IMAGE_ASPECT_COLOR_BIT
);
4949 /* Multi-planar images are not supported as attachments */
4950 assert(image
->aspects
== VK_IMAGE_ASPECT_COLOR_BIT
);
4951 assert(image
->n_planes
== 1);
4953 uint32_t base_clear_layer
= iview
->planes
[0].isl
.base_array_layer
;
4954 uint32_t clear_layer_count
= fb
->layers
;
4956 if (att_state
->fast_clear
&&
4957 do_first_layer_clear(cmd_state
, att_state
)) {
4958 /* We only support fast-clears on the first layer */
4959 assert(iview
->planes
[0].isl
.base_level
== 0);
4960 assert(iview
->planes
[0].isl
.base_array_layer
== 0);
4962 union isl_color_value clear_color
= {};
4963 anv_clear_color_from_att_state(&clear_color
, att_state
, iview
);
4964 if (iview
->image
->samples
== 1) {
4965 anv_image_ccs_op(cmd_buffer
, image
,
4966 iview
->planes
[0].isl
.format
,
4967 VK_IMAGE_ASPECT_COLOR_BIT
,
4968 0, 0, 1, ISL_AUX_OP_FAST_CLEAR
,
4972 anv_image_mcs_op(cmd_buffer
, image
,
4973 iview
->planes
[0].isl
.format
,
4974 VK_IMAGE_ASPECT_COLOR_BIT
,
4975 0, 1, ISL_AUX_OP_FAST_CLEAR
,
4980 clear_layer_count
--;
4982 att_state
->pending_clear_views
&= ~1;
4984 if (att_state
->clear_color_is_zero
) {
4985 /* This image has the auxiliary buffer enabled. We can mark the
4986 * subresource as not needing a resolve because the clear color
4987 * will match what's in every RENDER_SURFACE_STATE object when
4988 * it's being used for sampling.
4990 set_image_fast_clear_state(cmd_buffer
, iview
->image
,
4991 VK_IMAGE_ASPECT_COLOR_BIT
,
4992 ANV_FAST_CLEAR_DEFAULT_VALUE
);
4994 set_image_fast_clear_state(cmd_buffer
, iview
->image
,
4995 VK_IMAGE_ASPECT_COLOR_BIT
,
4996 ANV_FAST_CLEAR_ANY
);
5000 /* From the VkFramebufferCreateInfo spec:
5002 * "If the render pass uses multiview, then layers must be one and each
5003 * attachment requires a number of layers that is greater than the
5004 * maximum bit index set in the view mask in the subpasses in which it
5007 * So if multiview is active we ignore the number of layers in the
5008 * framebuffer and instead we honor the view mask from the subpass.
5011 assert(image
->n_planes
== 1);
5012 uint32_t pending_clear_mask
=
5013 get_multiview_subpass_clear_mask(cmd_state
, att_state
);
5016 for_each_bit(layer_idx
, pending_clear_mask
) {
5018 iview
->planes
[0].isl
.base_array_layer
+ layer_idx
;
5020 anv_image_clear_color(cmd_buffer
, image
,
5021 VK_IMAGE_ASPECT_COLOR_BIT
,
5022 att_state
->aux_usage
,
5023 iview
->planes
[0].isl
.format
,
5024 iview
->planes
[0].isl
.swizzle
,
5025 iview
->planes
[0].isl
.base_level
,
5028 vk_to_isl_color(att_state
->clear_value
.color
));
5031 att_state
->pending_clear_views
&= ~pending_clear_mask
;
5032 } else if (clear_layer_count
> 0) {
5033 assert(image
->n_planes
== 1);
5034 anv_image_clear_color(cmd_buffer
, image
, VK_IMAGE_ASPECT_COLOR_BIT
,
5035 att_state
->aux_usage
,
5036 iview
->planes
[0].isl
.format
,
5037 iview
->planes
[0].isl
.swizzle
,
5038 iview
->planes
[0].isl
.base_level
,
5039 base_clear_layer
, clear_layer_count
,
5041 vk_to_isl_color(att_state
->clear_value
.color
));
5043 } else if (att_state
->pending_clear_aspects
& (VK_IMAGE_ASPECT_DEPTH_BIT
|
5044 VK_IMAGE_ASPECT_STENCIL_BIT
)) {
5045 if (att_state
->fast_clear
&& !is_multiview
) {
5046 /* We currently only support HiZ for single-layer images */
5047 if (att_state
->pending_clear_aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
) {
5048 assert(iview
->image
->planes
[0].aux_usage
== ISL_AUX_USAGE_HIZ
);
5049 assert(iview
->planes
[0].isl
.base_level
== 0);
5050 assert(iview
->planes
[0].isl
.base_array_layer
== 0);
5051 assert(fb
->layers
== 1);
5054 anv_image_hiz_clear(cmd_buffer
, image
,
5055 att_state
->pending_clear_aspects
,
5056 iview
->planes
[0].isl
.base_level
,
5057 iview
->planes
[0].isl
.base_array_layer
,
5058 fb
->layers
, render_area
,
5059 att_state
->clear_value
.depthStencil
.stencil
);
5060 } else if (is_multiview
) {
5061 uint32_t pending_clear_mask
=
5062 get_multiview_subpass_clear_mask(cmd_state
, att_state
);
5065 for_each_bit(layer_idx
, pending_clear_mask
) {
5067 iview
->planes
[0].isl
.base_array_layer
+ layer_idx
;
5069 anv_image_clear_depth_stencil(cmd_buffer
, image
,
5070 att_state
->pending_clear_aspects
,
5071 att_state
->aux_usage
,
5072 iview
->planes
[0].isl
.base_level
,
5075 att_state
->clear_value
.depthStencil
.depth
,
5076 att_state
->clear_value
.depthStencil
.stencil
);
5079 att_state
->pending_clear_views
&= ~pending_clear_mask
;
5081 anv_image_clear_depth_stencil(cmd_buffer
, image
,
5082 att_state
->pending_clear_aspects
,
5083 att_state
->aux_usage
,
5084 iview
->planes
[0].isl
.base_level
,
5085 iview
->planes
[0].isl
.base_array_layer
,
5086 fb
->layers
, render_area
,
5087 att_state
->clear_value
.depthStencil
.depth
,
5088 att_state
->clear_value
.depthStencil
.stencil
);
5091 assert(att_state
->pending_clear_aspects
== 0);
5095 (att_state
->pending_load_aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) &&
5096 image
->planes
[0].aux_usage
!= ISL_AUX_USAGE_NONE
&&
5097 iview
->planes
[0].isl
.base_level
== 0 &&
5098 iview
->planes
[0].isl
.base_array_layer
== 0) {
5099 if (att_state
->aux_usage
!= ISL_AUX_USAGE_NONE
) {
5100 genX(copy_fast_clear_dwords
)(cmd_buffer
, att_state
->color
.state
,
5101 image
, VK_IMAGE_ASPECT_COLOR_BIT
,
5102 false /* copy to ss */);
5105 if (need_input_attachment_state(&cmd_state
->pass
->attachments
[a
]) &&
5106 att_state
->input_aux_usage
!= ISL_AUX_USAGE_NONE
) {
5107 genX(copy_fast_clear_dwords
)(cmd_buffer
, att_state
->input
.state
,
5108 image
, VK_IMAGE_ASPECT_COLOR_BIT
,
5109 false /* copy to ss */);
5113 if (subpass
->attachments
[i
].usage
==
5114 VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT
) {
5115 /* We assume that if we're starting a subpass, we're going to do some
5116 * rendering so we may end up with compressed data.
5118 genX(cmd_buffer_mark_image_written
)(cmd_buffer
, iview
->image
,
5119 VK_IMAGE_ASPECT_COLOR_BIT
,
5120 att_state
->aux_usage
,
5121 iview
->planes
[0].isl
.base_level
,
5122 iview
->planes
[0].isl
.base_array_layer
,
5124 } else if (subpass
->attachments
[i
].usage
==
5125 VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT
) {
5126 /* We may be writing depth or stencil so we need to mark the surface.
5127 * Unfortunately, there's no way to know at this point whether the
5128 * depth or stencil tests used will actually write to the surface.
5130 * Even though stencil may be plane 1, it always shares a base_level
5133 const struct isl_view
*ds_view
= &iview
->planes
[0].isl
;
5134 if (iview
->aspect_mask
& VK_IMAGE_ASPECT_DEPTH_BIT
) {
5135 genX(cmd_buffer_mark_image_written
)(cmd_buffer
, image
,
5136 VK_IMAGE_ASPECT_DEPTH_BIT
,
5137 att_state
->aux_usage
,
5138 ds_view
->base_level
,
5139 ds_view
->base_array_layer
,
5142 if (iview
->aspect_mask
& VK_IMAGE_ASPECT_STENCIL_BIT
) {
5143 /* Even though stencil may be plane 1, it always shares a
5144 * base_level with depth.
5146 genX(cmd_buffer_mark_image_written
)(cmd_buffer
, image
,
5147 VK_IMAGE_ASPECT_STENCIL_BIT
,
5149 ds_view
->base_level
,
5150 ds_view
->base_array_layer
,
5155 /* If multiview is enabled, then we are only done clearing when we no
5156 * longer have pending layers to clear, or when we have processed the
5157 * last subpass that uses this attachment.
5159 if (!is_multiview
||
5160 att_state
->pending_clear_views
== 0 ||
5161 current_subpass_is_last_for_attachment(cmd_state
, a
)) {
5162 att_state
->pending_clear_aspects
= 0;
5165 att_state
->pending_load_aspects
= 0;
5169 /* The PIPE_CONTROL command description says:
5171 * "Whenever a Binding Table Index (BTI) used by a Render Taget Message
5172 * points to a different RENDER_SURFACE_STATE, SW must issue a Render
5173 * Target Cache Flush by enabling this bit. When render target flush
5174 * is set due to new association of BTI, PS Scoreboard Stall bit must
5175 * be set in this packet."
5177 cmd_buffer
->state
.pending_pipe_bits
|=
5178 ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT
|
5179 ANV_PIPE_STALL_AT_SCOREBOARD_BIT
;
5183 /* GEN:BUG:14010455700
5185 * ISL will change some CHICKEN registers depending on the depth surface
5186 * format, along with emitting the depth and stencil packets. In that case,
5187 * we want to do a depth flush and stall, so the pipeline is not using these
5188 * settings while we change the registers.
5190 cmd_buffer
->state
.pending_pipe_bits
|=
5191 ANV_PIPE_DEPTH_CACHE_FLUSH_BIT
|
5192 ANV_PIPE_DEPTH_STALL_BIT
|
5193 ANV_PIPE_END_OF_PIPE_SYNC_BIT
;
5194 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
5197 cmd_buffer_emit_depth_stencil(cmd_buffer
);
5200 static enum blorp_filter
5201 vk_to_blorp_resolve_mode(VkResolveModeFlagBitsKHR vk_mode
)
5204 case VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
:
5205 return BLORP_FILTER_SAMPLE_0
;
5206 case VK_RESOLVE_MODE_AVERAGE_BIT_KHR
:
5207 return BLORP_FILTER_AVERAGE
;
5208 case VK_RESOLVE_MODE_MIN_BIT_KHR
:
5209 return BLORP_FILTER_MIN_SAMPLE
;
5210 case VK_RESOLVE_MODE_MAX_BIT_KHR
:
5211 return BLORP_FILTER_MAX_SAMPLE
;
5213 return BLORP_FILTER_NONE
;
5218 cmd_buffer_end_subpass(struct anv_cmd_buffer
*cmd_buffer
)
5220 struct anv_cmd_state
*cmd_state
= &cmd_buffer
->state
;
5221 struct anv_subpass
*subpass
= cmd_state
->subpass
;
5222 uint32_t subpass_id
= anv_get_subpass_id(&cmd_buffer
->state
);
5223 struct anv_framebuffer
*fb
= cmd_buffer
->state
.framebuffer
;
5225 if (subpass
->has_color_resolve
) {
5226 /* We are about to do some MSAA resolves. We need to flush so that the
5227 * result of writes to the MSAA color attachments show up in the sampler
5228 * when we blit to the single-sampled resolve target.
5230 cmd_buffer
->state
.pending_pipe_bits
|=
5231 ANV_PIPE_TEXTURE_CACHE_INVALIDATE_BIT
|
5232 ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT
;
5234 for (uint32_t i
= 0; i
< subpass
->color_count
; ++i
) {
5235 uint32_t src_att
= subpass
->color_attachments
[i
].attachment
;
5236 uint32_t dst_att
= subpass
->resolve_attachments
[i
].attachment
;
5238 if (dst_att
== VK_ATTACHMENT_UNUSED
)
5241 assert(src_att
< cmd_buffer
->state
.pass
->attachment_count
);
5242 assert(dst_att
< cmd_buffer
->state
.pass
->attachment_count
);
5244 if (cmd_buffer
->state
.attachments
[dst_att
].pending_clear_aspects
) {
5245 /* From the Vulkan 1.0 spec:
5247 * If the first use of an attachment in a render pass is as a
5248 * resolve attachment, then the loadOp is effectively ignored
5249 * as the resolve is guaranteed to overwrite all pixels in the
5252 cmd_buffer
->state
.attachments
[dst_att
].pending_clear_aspects
= 0;
5255 struct anv_image_view
*src_iview
= cmd_state
->attachments
[src_att
].image_view
;
5256 struct anv_image_view
*dst_iview
= cmd_state
->attachments
[dst_att
].image_view
;
5258 const VkRect2D render_area
= cmd_buffer
->state
.render_area
;
5260 enum isl_aux_usage src_aux_usage
=
5261 cmd_buffer
->state
.attachments
[src_att
].aux_usage
;
5262 enum isl_aux_usage dst_aux_usage
=
5263 cmd_buffer
->state
.attachments
[dst_att
].aux_usage
;
5265 assert(src_iview
->aspect_mask
== VK_IMAGE_ASPECT_COLOR_BIT
&&
5266 dst_iview
->aspect_mask
== VK_IMAGE_ASPECT_COLOR_BIT
);
5268 anv_image_msaa_resolve(cmd_buffer
,
5269 src_iview
->image
, src_aux_usage
,
5270 src_iview
->planes
[0].isl
.base_level
,
5271 src_iview
->planes
[0].isl
.base_array_layer
,
5272 dst_iview
->image
, dst_aux_usage
,
5273 dst_iview
->planes
[0].isl
.base_level
,
5274 dst_iview
->planes
[0].isl
.base_array_layer
,
5275 VK_IMAGE_ASPECT_COLOR_BIT
,
5276 render_area
.offset
.x
, render_area
.offset
.y
,
5277 render_area
.offset
.x
, render_area
.offset
.y
,
5278 render_area
.extent
.width
,
5279 render_area
.extent
.height
,
5280 fb
->layers
, BLORP_FILTER_NONE
);
5284 if (subpass
->ds_resolve_attachment
) {
5285 /* We are about to do some MSAA resolves. We need to flush so that the
5286 * result of writes to the MSAA depth attachments show up in the sampler
5287 * when we blit to the single-sampled resolve target.
5289 cmd_buffer
->state
.pending_pipe_bits
|=
5290 ANV_PIPE_TEXTURE_CACHE_INVALIDATE_BIT
|
5291 ANV_PIPE_DEPTH_CACHE_FLUSH_BIT
;
5293 uint32_t src_att
= subpass
->depth_stencil_attachment
->attachment
;
5294 uint32_t dst_att
= subpass
->ds_resolve_attachment
->attachment
;
5296 assert(src_att
< cmd_buffer
->state
.pass
->attachment_count
);
5297 assert(dst_att
< cmd_buffer
->state
.pass
->attachment_count
);
5299 if (cmd_buffer
->state
.attachments
[dst_att
].pending_clear_aspects
) {
5300 /* From the Vulkan 1.0 spec:
5302 * If the first use of an attachment in a render pass is as a
5303 * resolve attachment, then the loadOp is effectively ignored
5304 * as the resolve is guaranteed to overwrite all pixels in the
5307 cmd_buffer
->state
.attachments
[dst_att
].pending_clear_aspects
= 0;
5310 struct anv_image_view
*src_iview
= cmd_state
->attachments
[src_att
].image_view
;
5311 struct anv_image_view
*dst_iview
= cmd_state
->attachments
[dst_att
].image_view
;
5313 const VkRect2D render_area
= cmd_buffer
->state
.render_area
;
5315 struct anv_attachment_state
*src_state
=
5316 &cmd_state
->attachments
[src_att
];
5317 struct anv_attachment_state
*dst_state
=
5318 &cmd_state
->attachments
[dst_att
];
5320 if ((src_iview
->image
->aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
) &&
5321 subpass
->depth_resolve_mode
!= VK_RESOLVE_MODE_NONE_KHR
) {
5323 /* MSAA resolves sample from the source attachment. Transition the
5324 * depth attachment first to get rid of any HiZ that we may not be
5327 transition_depth_buffer(cmd_buffer
, src_iview
->image
,
5328 src_state
->current_layout
,
5329 VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL
);
5330 src_state
->aux_usage
=
5331 anv_layout_to_aux_usage(&cmd_buffer
->device
->info
, src_iview
->image
,
5332 VK_IMAGE_ASPECT_DEPTH_BIT
,
5333 VK_IMAGE_USAGE_TRANSFER_SRC_BIT
,
5334 VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL
);
5335 src_state
->current_layout
= VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL
;
5337 /* MSAA resolves write to the resolve attachment as if it were any
5338 * other transfer op. Transition the resolve attachment accordingly.
5340 VkImageLayout dst_initial_layout
= dst_state
->current_layout
;
5342 /* If our render area is the entire size of the image, we're going to
5343 * blow it all away so we can claim the initial layout is UNDEFINED
5344 * and we'll get a HiZ ambiguate instead of a resolve.
5346 if (dst_iview
->image
->type
!= VK_IMAGE_TYPE_3D
&&
5347 render_area
.offset
.x
== 0 && render_area
.offset
.y
== 0 &&
5348 render_area
.extent
.width
== dst_iview
->extent
.width
&&
5349 render_area
.extent
.height
== dst_iview
->extent
.height
)
5350 dst_initial_layout
= VK_IMAGE_LAYOUT_UNDEFINED
;
5352 transition_depth_buffer(cmd_buffer
, dst_iview
->image
,
5354 VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
);
5355 dst_state
->aux_usage
=
5356 anv_layout_to_aux_usage(&cmd_buffer
->device
->info
, dst_iview
->image
,
5357 VK_IMAGE_ASPECT_DEPTH_BIT
,
5358 VK_IMAGE_USAGE_TRANSFER_DST_BIT
,
5359 VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
);
5360 dst_state
->current_layout
= VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
;
5362 enum blorp_filter filter
=
5363 vk_to_blorp_resolve_mode(subpass
->depth_resolve_mode
);
5365 anv_image_msaa_resolve(cmd_buffer
,
5366 src_iview
->image
, src_state
->aux_usage
,
5367 src_iview
->planes
[0].isl
.base_level
,
5368 src_iview
->planes
[0].isl
.base_array_layer
,
5369 dst_iview
->image
, dst_state
->aux_usage
,
5370 dst_iview
->planes
[0].isl
.base_level
,
5371 dst_iview
->planes
[0].isl
.base_array_layer
,
5372 VK_IMAGE_ASPECT_DEPTH_BIT
,
5373 render_area
.offset
.x
, render_area
.offset
.y
,
5374 render_area
.offset
.x
, render_area
.offset
.y
,
5375 render_area
.extent
.width
,
5376 render_area
.extent
.height
,
5377 fb
->layers
, filter
);
5380 if ((src_iview
->image
->aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
) &&
5381 subpass
->stencil_resolve_mode
!= VK_RESOLVE_MODE_NONE_KHR
) {
5383 src_state
->current_stencil_layout
= VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL
;
5384 dst_state
->current_stencil_layout
= VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
;
5386 enum isl_aux_usage src_aux_usage
= ISL_AUX_USAGE_NONE
;
5387 enum isl_aux_usage dst_aux_usage
= ISL_AUX_USAGE_NONE
;
5389 enum blorp_filter filter
=
5390 vk_to_blorp_resolve_mode(subpass
->stencil_resolve_mode
);
5392 anv_image_msaa_resolve(cmd_buffer
,
5393 src_iview
->image
, src_aux_usage
,
5394 src_iview
->planes
[0].isl
.base_level
,
5395 src_iview
->planes
[0].isl
.base_array_layer
,
5396 dst_iview
->image
, dst_aux_usage
,
5397 dst_iview
->planes
[0].isl
.base_level
,
5398 dst_iview
->planes
[0].isl
.base_array_layer
,
5399 VK_IMAGE_ASPECT_STENCIL_BIT
,
5400 render_area
.offset
.x
, render_area
.offset
.y
,
5401 render_area
.offset
.x
, render_area
.offset
.y
,
5402 render_area
.extent
.width
,
5403 render_area
.extent
.height
,
5404 fb
->layers
, filter
);
5409 /* On gen7, we have to store a texturable version of the stencil buffer in
5410 * a shadow whenever VK_IMAGE_USAGE_SAMPLED_BIT is set and copy back and
5411 * forth at strategic points. Stencil writes are only allowed in following
5414 * - VK_IMAGE_LAYOUT_GENERAL
5415 * - VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
5416 * - VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL
5417 * - VK_IMAGE_LAYOUT_DEPTH_READ_ONLY_STENCIL_ATTACHMENT_OPTIMAL
5418 * - VK_IMAGE_LAYOUT_STENCIL_ATTACHMENT_OPTIMAL_KHR
5420 * For general, we have no nice opportunity to transition so we do the copy
5421 * to the shadow unconditionally at the end of the subpass. For transfer
5422 * destinations, we can update it as part of the transfer op. For the other
5423 * layouts, we delay the copy until a transition into some other layout.
5425 if (subpass
->depth_stencil_attachment
) {
5426 uint32_t a
= subpass
->depth_stencil_attachment
->attachment
;
5427 assert(a
!= VK_ATTACHMENT_UNUSED
);
5429 struct anv_attachment_state
*att_state
= &cmd_state
->attachments
[a
];
5430 struct anv_image_view
*iview
= cmd_state
->attachments
[a
].image_view
;;
5431 const struct anv_image
*image
= iview
->image
;
5433 if (image
->aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
) {
5434 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
5435 VK_IMAGE_ASPECT_STENCIL_BIT
);
5437 if (image
->planes
[plane
].shadow_surface
.isl
.size_B
> 0 &&
5438 att_state
->current_stencil_layout
== VK_IMAGE_LAYOUT_GENERAL
) {
5439 assert(image
->aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
);
5440 anv_image_copy_to_shadow(cmd_buffer
, image
,
5441 VK_IMAGE_ASPECT_STENCIL_BIT
,
5442 iview
->planes
[plane
].isl
.base_level
, 1,
5443 iview
->planes
[plane
].isl
.base_array_layer
,
5448 #endif /* GEN_GEN == 7 */
5450 for (uint32_t i
= 0; i
< subpass
->attachment_count
; ++i
) {
5451 const uint32_t a
= subpass
->attachments
[i
].attachment
;
5452 if (a
== VK_ATTACHMENT_UNUSED
)
5455 if (cmd_state
->pass
->attachments
[a
].last_subpass_idx
!= subpass_id
)
5458 assert(a
< cmd_state
->pass
->attachment_count
);
5459 struct anv_attachment_state
*att_state
= &cmd_state
->attachments
[a
];
5460 struct anv_image_view
*iview
= cmd_state
->attachments
[a
].image_view
;
5461 const struct anv_image
*image
= iview
->image
;
5463 if ((image
->aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) &&
5464 image
->vk_format
!= iview
->vk_format
) {
5465 enum anv_fast_clear_type fast_clear_type
=
5466 anv_layout_to_fast_clear_type(&cmd_buffer
->device
->info
,
5467 image
, VK_IMAGE_ASPECT_COLOR_BIT
,
5468 att_state
->current_layout
);
5470 /* If any clear color was used, flush it down the aux surfaces. If we
5471 * don't do it now using the view's format we might use the clear
5472 * color incorrectly in the following resolves (for example with an
5473 * SRGB view & a UNORM image).
5475 if (fast_clear_type
!= ANV_FAST_CLEAR_NONE
) {
5476 anv_perf_warn(cmd_buffer
->device
, iview
,
5477 "Doing a partial resolve to get rid of clear color at the "
5478 "end of a renderpass due to an image/view format mismatch");
5480 uint32_t base_layer
, layer_count
;
5481 if (image
->type
== VK_IMAGE_TYPE_3D
) {
5483 layer_count
= anv_minify(iview
->image
->extent
.depth
,
5484 iview
->planes
[0].isl
.base_level
);
5486 base_layer
= iview
->planes
[0].isl
.base_array_layer
;
5487 layer_count
= fb
->layers
;
5490 for (uint32_t a
= 0; a
< layer_count
; a
++) {
5491 uint32_t array_layer
= base_layer
+ a
;
5492 if (image
->samples
== 1) {
5493 anv_cmd_predicated_ccs_resolve(cmd_buffer
, image
,
5494 iview
->planes
[0].isl
.format
,
5495 VK_IMAGE_ASPECT_COLOR_BIT
,
5496 iview
->planes
[0].isl
.base_level
,
5498 ISL_AUX_OP_PARTIAL_RESOLVE
,
5499 ANV_FAST_CLEAR_NONE
);
5501 anv_cmd_predicated_mcs_resolve(cmd_buffer
, image
,
5502 iview
->planes
[0].isl
.format
,
5503 VK_IMAGE_ASPECT_COLOR_BIT
,
5505 ISL_AUX_OP_PARTIAL_RESOLVE
,
5506 ANV_FAST_CLEAR_NONE
);
5512 /* Transition the image into the final layout for this render pass */
5513 VkImageLayout target_layout
=
5514 cmd_state
->pass
->attachments
[a
].final_layout
;
5515 VkImageLayout target_stencil_layout
=
5516 cmd_state
->pass
->attachments
[a
].stencil_final_layout
;
5518 uint32_t base_layer
, layer_count
;
5519 if (image
->type
== VK_IMAGE_TYPE_3D
) {
5521 layer_count
= anv_minify(iview
->image
->extent
.depth
,
5522 iview
->planes
[0].isl
.base_level
);
5524 base_layer
= iview
->planes
[0].isl
.base_array_layer
;
5525 layer_count
= fb
->layers
;
5528 if (image
->aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) {
5529 assert(image
->aspects
== VK_IMAGE_ASPECT_COLOR_BIT
);
5530 transition_color_buffer(cmd_buffer
, image
, VK_IMAGE_ASPECT_COLOR_BIT
,
5531 iview
->planes
[0].isl
.base_level
, 1,
5532 base_layer
, layer_count
,
5533 att_state
->current_layout
, target_layout
);
5536 if (image
->aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
) {
5537 transition_depth_buffer(cmd_buffer
, image
,
5538 att_state
->current_layout
, target_layout
);
5541 if (image
->aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
) {
5542 transition_stencil_buffer(cmd_buffer
, image
,
5543 iview
->planes
[0].isl
.base_level
, 1,
5544 base_layer
, layer_count
,
5545 att_state
->current_stencil_layout
,
5546 target_stencil_layout
);
5550 /* Accumulate any subpass flushes that need to happen after the subpass.
5551 * Yes, they do get accumulated twice in the NextSubpass case but since
5552 * genX_CmdNextSubpass just calls end/begin back-to-back, we just end up
5553 * ORing the bits in twice so it's harmless.
5555 cmd_buffer
->state
.pending_pipe_bits
|=
5556 cmd_buffer
->state
.pass
->subpass_flushes
[subpass_id
+ 1];
5559 void genX(CmdBeginRenderPass
)(
5560 VkCommandBuffer commandBuffer
,
5561 const VkRenderPassBeginInfo
* pRenderPassBegin
,
5562 VkSubpassContents contents
)
5564 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
5565 ANV_FROM_HANDLE(anv_render_pass
, pass
, pRenderPassBegin
->renderPass
);
5566 ANV_FROM_HANDLE(anv_framebuffer
, framebuffer
, pRenderPassBegin
->framebuffer
);
5568 cmd_buffer
->state
.framebuffer
= framebuffer
;
5569 cmd_buffer
->state
.pass
= pass
;
5570 cmd_buffer
->state
.render_area
= pRenderPassBegin
->renderArea
;
5572 genX(cmd_buffer_setup_attachments
)(cmd_buffer
, pass
, pRenderPassBegin
);
5574 /* If we failed to setup the attachments we should not try to go further */
5575 if (result
!= VK_SUCCESS
) {
5576 assert(anv_batch_has_error(&cmd_buffer
->batch
));
5580 genX(flush_pipeline_select_3d
)(cmd_buffer
);
5582 cmd_buffer_begin_subpass(cmd_buffer
, 0);
5585 void genX(CmdBeginRenderPass2
)(
5586 VkCommandBuffer commandBuffer
,
5587 const VkRenderPassBeginInfo
* pRenderPassBeginInfo
,
5588 const VkSubpassBeginInfoKHR
* pSubpassBeginInfo
)
5590 genX(CmdBeginRenderPass
)(commandBuffer
, pRenderPassBeginInfo
,
5591 pSubpassBeginInfo
->contents
);
5594 void genX(CmdNextSubpass
)(
5595 VkCommandBuffer commandBuffer
,
5596 VkSubpassContents contents
)
5598 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
5600 if (anv_batch_has_error(&cmd_buffer
->batch
))
5603 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
5605 uint32_t prev_subpass
= anv_get_subpass_id(&cmd_buffer
->state
);
5606 cmd_buffer_end_subpass(cmd_buffer
);
5607 cmd_buffer_begin_subpass(cmd_buffer
, prev_subpass
+ 1);
5610 void genX(CmdNextSubpass2
)(
5611 VkCommandBuffer commandBuffer
,
5612 const VkSubpassBeginInfoKHR
* pSubpassBeginInfo
,
5613 const VkSubpassEndInfoKHR
* pSubpassEndInfo
)
5615 genX(CmdNextSubpass
)(commandBuffer
, pSubpassBeginInfo
->contents
);
5618 void genX(CmdEndRenderPass
)(
5619 VkCommandBuffer commandBuffer
)
5621 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
5623 if (anv_batch_has_error(&cmd_buffer
->batch
))
5626 cmd_buffer_end_subpass(cmd_buffer
);
5628 cmd_buffer
->state
.hiz_enabled
= false;
5631 anv_dump_add_attachments(cmd_buffer
);
5634 /* Remove references to render pass specific state. This enables us to
5635 * detect whether or not we're in a renderpass.
5637 cmd_buffer
->state
.framebuffer
= NULL
;
5638 cmd_buffer
->state
.pass
= NULL
;
5639 cmd_buffer
->state
.subpass
= NULL
;
5642 void genX(CmdEndRenderPass2
)(
5643 VkCommandBuffer commandBuffer
,
5644 const VkSubpassEndInfoKHR
* pSubpassEndInfo
)
5646 genX(CmdEndRenderPass
)(commandBuffer
);
5650 genX(cmd_emit_conditional_render_predicate
)(struct anv_cmd_buffer
*cmd_buffer
)
5652 #if GEN_GEN >= 8 || GEN_IS_HASWELL
5653 struct gen_mi_builder b
;
5654 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
5656 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC0
),
5657 gen_mi_reg32(ANV_PREDICATE_RESULT_REG
));
5658 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC1
), gen_mi_imm(0));
5660 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_PREDICATE
), mip
) {
5661 mip
.LoadOperation
= LOAD_LOADINV
;
5662 mip
.CombineOperation
= COMBINE_SET
;
5663 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
5668 #if GEN_GEN >= 8 || GEN_IS_HASWELL
5669 void genX(CmdBeginConditionalRenderingEXT
)(
5670 VkCommandBuffer commandBuffer
,
5671 const VkConditionalRenderingBeginInfoEXT
* pConditionalRenderingBegin
)
5673 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
5674 ANV_FROM_HANDLE(anv_buffer
, buffer
, pConditionalRenderingBegin
->buffer
);
5675 struct anv_cmd_state
*cmd_state
= &cmd_buffer
->state
;
5676 struct anv_address value_address
=
5677 anv_address_add(buffer
->address
, pConditionalRenderingBegin
->offset
);
5679 const bool isInverted
= pConditionalRenderingBegin
->flags
&
5680 VK_CONDITIONAL_RENDERING_INVERTED_BIT_EXT
;
5682 cmd_state
->conditional_render_enabled
= true;
5684 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
5686 struct gen_mi_builder b
;
5687 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
5689 /* Section 19.4 of the Vulkan 1.1.85 spec says:
5691 * If the value of the predicate in buffer memory changes
5692 * while conditional rendering is active, the rendering commands
5693 * may be discarded in an implementation-dependent way.
5694 * Some implementations may latch the value of the predicate
5695 * upon beginning conditional rendering while others
5696 * may read it before every rendering command.
5698 * So it's perfectly fine to read a value from the buffer once.
5700 struct gen_mi_value value
= gen_mi_mem32(value_address
);
5702 /* Precompute predicate result, it is necessary to support secondary
5703 * command buffers since it is unknown if conditional rendering is
5704 * inverted when populating them.
5706 gen_mi_store(&b
, gen_mi_reg64(ANV_PREDICATE_RESULT_REG
),
5707 isInverted
? gen_mi_uge(&b
, gen_mi_imm(0), value
) :
5708 gen_mi_ult(&b
, gen_mi_imm(0), value
));
5711 void genX(CmdEndConditionalRenderingEXT
)(
5712 VkCommandBuffer commandBuffer
)
5714 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
5715 struct anv_cmd_state
*cmd_state
= &cmd_buffer
->state
;
5717 cmd_state
->conditional_render_enabled
= false;
5721 /* Set of stage bits for which are pipelined, i.e. they get queued by the
5722 * command streamer for later execution.
5724 #define ANV_PIPELINE_STAGE_PIPELINED_BITS \
5725 (VK_PIPELINE_STAGE_VERTEX_INPUT_BIT | \
5726 VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | \
5727 VK_PIPELINE_STAGE_TESSELLATION_CONTROL_SHADER_BIT | \
5728 VK_PIPELINE_STAGE_TESSELLATION_EVALUATION_SHADER_BIT | \
5729 VK_PIPELINE_STAGE_GEOMETRY_SHADER_BIT | \
5730 VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | \
5731 VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT | \
5732 VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT | \
5733 VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT | \
5734 VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT | \
5735 VK_PIPELINE_STAGE_TRANSFER_BIT | \
5736 VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT | \
5737 VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT | \
5738 VK_PIPELINE_STAGE_ALL_COMMANDS_BIT)
5740 void genX(CmdSetEvent
)(
5741 VkCommandBuffer commandBuffer
,
5743 VkPipelineStageFlags stageMask
)
5745 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
5746 ANV_FROM_HANDLE(anv_event
, event
, _event
);
5748 cmd_buffer
->state
.pending_pipe_bits
|= ANV_PIPE_POST_SYNC_BIT
;
5749 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
5751 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
5752 if (stageMask
& ANV_PIPELINE_STAGE_PIPELINED_BITS
) {
5753 pc
.StallAtPixelScoreboard
= true;
5754 pc
.CommandStreamerStallEnable
= true;
5757 pc
.DestinationAddressType
= DAT_PPGTT
,
5758 pc
.PostSyncOperation
= WriteImmediateData
,
5759 pc
.Address
= (struct anv_address
) {
5760 cmd_buffer
->device
->dynamic_state_pool
.block_pool
.bo
,
5763 pc
.ImmediateData
= VK_EVENT_SET
;
5767 void genX(CmdResetEvent
)(
5768 VkCommandBuffer commandBuffer
,
5770 VkPipelineStageFlags stageMask
)
5772 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
5773 ANV_FROM_HANDLE(anv_event
, event
, _event
);
5775 cmd_buffer
->state
.pending_pipe_bits
|= ANV_PIPE_POST_SYNC_BIT
;
5776 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
5778 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
5779 if (stageMask
& ANV_PIPELINE_STAGE_PIPELINED_BITS
) {
5780 pc
.StallAtPixelScoreboard
= true;
5781 pc
.CommandStreamerStallEnable
= true;
5784 pc
.DestinationAddressType
= DAT_PPGTT
;
5785 pc
.PostSyncOperation
= WriteImmediateData
;
5786 pc
.Address
= (struct anv_address
) {
5787 cmd_buffer
->device
->dynamic_state_pool
.block_pool
.bo
,
5790 pc
.ImmediateData
= VK_EVENT_RESET
;
5794 void genX(CmdWaitEvents
)(
5795 VkCommandBuffer commandBuffer
,
5796 uint32_t eventCount
,
5797 const VkEvent
* pEvents
,
5798 VkPipelineStageFlags srcStageMask
,
5799 VkPipelineStageFlags destStageMask
,
5800 uint32_t memoryBarrierCount
,
5801 const VkMemoryBarrier
* pMemoryBarriers
,
5802 uint32_t bufferMemoryBarrierCount
,
5803 const VkBufferMemoryBarrier
* pBufferMemoryBarriers
,
5804 uint32_t imageMemoryBarrierCount
,
5805 const VkImageMemoryBarrier
* pImageMemoryBarriers
)
5808 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
5810 for (uint32_t i
= 0; i
< eventCount
; i
++) {
5811 ANV_FROM_HANDLE(anv_event
, event
, pEvents
[i
]);
5813 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_SEMAPHORE_WAIT
), sem
) {
5814 sem
.WaitMode
= PollingMode
,
5815 sem
.CompareOperation
= COMPARE_SAD_EQUAL_SDD
,
5816 sem
.SemaphoreDataDword
= VK_EVENT_SET
,
5817 sem
.SemaphoreAddress
= (struct anv_address
) {
5818 cmd_buffer
->device
->dynamic_state_pool
.block_pool
.bo
,
5824 anv_finishme("Implement events on gen7");
5827 genX(CmdPipelineBarrier
)(commandBuffer
, srcStageMask
, destStageMask
,
5828 false, /* byRegion */
5829 memoryBarrierCount
, pMemoryBarriers
,
5830 bufferMemoryBarrierCount
, pBufferMemoryBarriers
,
5831 imageMemoryBarrierCount
, pImageMemoryBarriers
);
5834 VkResult
genX(CmdSetPerformanceOverrideINTEL
)(
5835 VkCommandBuffer commandBuffer
,
5836 const VkPerformanceOverrideInfoINTEL
* pOverrideInfo
)
5838 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
5840 switch (pOverrideInfo
->type
) {
5841 case VK_PERFORMANCE_OVERRIDE_TYPE_NULL_HARDWARE_INTEL
: {
5845 anv_pack_struct(&dw
, GENX(CS_DEBUG_MODE2
),
5846 ._3DRenderingInstructionDisable
= pOverrideInfo
->enable
,
5847 .MediaInstructionDisable
= pOverrideInfo
->enable
,
5848 ._3DRenderingInstructionDisableMask
= true,
5849 .MediaInstructionDisableMask
= true);
5850 emit_lri(&cmd_buffer
->batch
, GENX(CS_DEBUG_MODE2_num
), dw
);
5852 anv_pack_struct(&dw
, GENX(INSTPM
),
5853 ._3DRenderingInstructionDisable
= pOverrideInfo
->enable
,
5854 .MediaInstructionDisable
= pOverrideInfo
->enable
,
5855 ._3DRenderingInstructionDisableMask
= true,
5856 .MediaInstructionDisableMask
= true);
5857 emit_lri(&cmd_buffer
->batch
, GENX(INSTPM_num
), dw
);
5862 case VK_PERFORMANCE_OVERRIDE_TYPE_FLUSH_GPU_CACHES_INTEL
:
5863 if (pOverrideInfo
->enable
) {
5864 /* FLUSH ALL THE THINGS! As requested by the MDAPI team. */
5865 cmd_buffer
->state
.pending_pipe_bits
|=
5866 ANV_PIPE_FLUSH_BITS
|
5867 ANV_PIPE_INVALIDATE_BITS
;
5868 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
5873 unreachable("Invalid override");
5879 VkResult
genX(CmdSetPerformanceStreamMarkerINTEL
)(
5880 VkCommandBuffer commandBuffer
,
5881 const VkPerformanceStreamMarkerInfoINTEL
* pMarkerInfo
)
5883 /* TODO: Waiting on the register to write, might depend on generation. */